<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"
	xmlns:content="http://purl.org/rss/1.0/modules/content/"
	xmlns:wfw="http://wellformedweb.org/CommentAPI/"
	xmlns:dc="http://purl.org/dc/elements/1.1/"
	xmlns:atom="http://www.w3.org/2005/Atom"
	xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
	xmlns:slash="http://purl.org/rss/1.0/modules/slash/"
	>

<channel>
	<title>artificial intelligence archivos | Fali Fuentes</title>
	<atom:link href="https://falifuentes.com/tag/artificial-intelligence/feed/" rel="self" type="application/rss+xml" />
	<link>https://falifuentes.com/tag/artificial-intelligence/</link>
	<description>Blog de Fali Fuentes (Málaga) &#124; Ciberseguridad, IA y Tecnología: Protege tu vida digital, domina tendencias tech y descubre análisis expertos.   ¡Actualizaciones diarias!</description>
	<lastBuildDate>Fri, 20 Nov 2026 05:12:08 +0000</lastBuildDate>
	<language>es</language>
	<sy:updatePeriod>
	hourly	</sy:updatePeriod>
	<sy:updateFrequency>
	1	</sy:updateFrequency>
	<generator>https://wordpress.org/?v=7.0</generator>

<image>
	<url>https://falifuentes.com/wp-content/uploads/2024/08/Favi_sec_p.png</url>
	<title>artificial intelligence archivos | Fali Fuentes</title>
	<link>https://falifuentes.com/tag/artificial-intelligence/</link>
	<width>32</width>
	<height>32</height>
</image> 
	<item>
		<title>Microsoft Ignite 2026: Copilot becomes the work operating system with Autopilot</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-20-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-20-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Fri, 20 Nov 2026 05:12:08 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Compliance and Regulations]]></category>
		<category><![CDATA[Email]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-20-am-en/</guid>

					<description><![CDATA[<p>On November 17, Microsoft launched Autopilot, Work IQ, and Agent 365 at Ignite, moving Copilot from assistant to execution and governance layer in your tenant.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-20-am-en/">Microsoft Ignite 2026: Copilot becomes the work operating system with Autopilot</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-20-am-es/" hreflang="es">Leer este artículo en español</a></p>
<p>On November 17, Microsoft wrapped up Ignite 2026 by shifting Copilot&#8217;s axis: it is no longer just a chat assistant, but the execution and governance layer that unifies models, agents, data, and applications inside your tenant. For you, who manage identity, security, or architecture in Microsoft 365 environments, the change is immediate: Autopilot begins rolling out in the Frontier program with its own identity (Entra Agent ID) and native containment via MXC, Intune, and Defender; Work IQ already reasons across 19.5 exabytes of email, Teams, and files; and Agent 365 adds «completed work» analytics and hard spending limits in Cowork. All of this lands today, not on the roadmap, according to the <a href='https://www.microsoft.com/en-us/copilot/blog/2026/11/17/copilot-at-microsoft-ignite-expanding-whats-possible-with-intelligence-built-for-work'>official Microsoft Copilot Blog announcement</a>.</p>
<h2>Autopilot launches in Frontier with its own identity and containment via MXC</h2>
<p>Autopilot isn&#8217;t just another bot: it lives in your tenant, has persistent memory, its own tools, and a designated human manager. Its deployment starts today for Frontier customers and forces you to review conditional access policies and audit agent actions. Local agent containment relies on Microsoft Execution Containers (MXC) integrated with Intune and Defender, but requires Windows 365 for Agents and Intune/Defender P2 licenses; some MCP/tool controls remain in preview. If you&#8217;re not in Frontier, plan the pilot with explicit access scope and don&#8217;t use it to replace human approvals for critical actions.</p>
<h2>Work IQ reasons across 19.5 exabytes and Workspaces enables multi-agent</h2>
<p>Work IQ is now available and knows what to search across 90 billion emails, 15 billion Teams messages, and 10 billion new files each week. Workspaces, also in Frontier, turns collaboration into true multiplayer: people and agents share context toward a common goal. For architects, this means the organization&#8217;s knowledge graph stops being passive and becomes an automatic decision surface; review sensitivity labels and DLP before exposing sensitive data to these flows.</p>
<h2>Copilot Managed Runtime is now GA and Copilot Code arrives in weeks</h2>
<p>Secure hosting for apps built in natural language (Copilot Managed Runtime) is GA as of today; Copilot Code, which lets non-developers build solutions, will be GA «in the coming weeks.» The apps inherit M365 permissions, sensitivity labels, and DLP, and are governed from the M365 Admin Center. Use it first for low-risk internal apps; validate that the runtime meets your data residency and isolation requirements before moving regulated workloads.</p>
<h2>On-demand secret scanning in Copilot CLI, app, and code review</h2>
<p>GitHub has enabled on-demand secret scanning inside Copilot&#8217;s security review in public preview: it detects over 500 credential types using provider patterns and a fine-tuned model that reads context (passwords in connection strings, etc.). You enable it manually per tool (CLI, app, SDK, code review) and it only analyzes the changes under review, not the repo history. Enable it now in your review pipelines; it reduces credential leakage in AI-generated code at no extra plan cost, according to the <a href='https://github.blog/changelog/2026-11-17-secret-scanning-on-demand-for-developers-and-agents'>GitHub changelog</a>.</p>
<h2>Cost and model governance: Business Analytics and Auto routing</h2>
<p>Agent 365 Business Analytics (Frontier Public) measures investment in «completed work hours,» not tokens, and Cost Management enforces hard per-team limits in Cowork (consumption in AI credits, Secret Protection SKU). Auto routing selects among GPT-6.2, Claude, and Microsoft models optimizing for quality, speed, and cost, but does not guarantee model-specific regulatory compliance; define model governance policies in the tenant and validate what data goes to each provider. The Salesforce integration (public preview, 38 skills) and Dynamics 365 (100+ skills) are managed in a unified catalog with central registration; configure CRM/ERP permissions before enabling.</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;Microsoft Ignite 2026: Copilot becomes the work operating system with Autopilot&quot;,
  &quot;summary&quot;: &quot;On November 17, Microsoft launched Autopilot, Work IQ, and Agent 365 at Ignite, moving Copilot from assistant to execution and governance layer in your tenant.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-20T06:01:16.075+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-17&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-20-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-20-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Compliance and Regulations&quot;,
    &quot;Email&quot;,
    &quot;English&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://www.microsoft.com/en-us/copilot/blog/2026/11/17/copilot-at-microsoft-ignite-expanding-whats-possible-with-intelligence-built-for-work&quot;,
    &quot;https://github.blog/changelog/2026-11-17-secret-scanning-on-demand-for-developers-and-agents&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-20-am-en/">Microsoft Ignite 2026: Copilot becomes the work operating system with Autopilot</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>AWS makes Bedrock AgentCore runtime V2 default and publishes self-healing ETL guide</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-18-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-18-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Wed, 18 Nov 2026 05:12:50 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[Big data]]></category>
		<category><![CDATA[Healthcare]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-18-am-en/</guid>

					<description><![CDATA[<p>Starting November 12, new AgentCore runtimes use V2 in 22 regions with <2 s startup and scale-to-zero; AWS publishes a self-healing ETL pattern with four agents and dual correction.
</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-18-am-en/">AWS makes Bedrock AgentCore runtime V2 default and publishes self-healing ETL guide</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-18-am-es/" hreflang="es">Leer este artículo en español</a></p>
<p>AWS has changed the default runtime of Amazon Bedrock AgentCore to version V2 in 22 regions —including Europe (Spain, Frankfurt, Ireland, London, Milan, Paris), the U.S., Canada, South America (São Paulo), and Asia Pacific— starting November 12, 2026. Existing runtimes remain on V1 unless you explicitly update them with <code>platformVersion=V2</code>; if you need to force V1 in a region where V2 is the default, use <code>platformVersion=V1</code> when creating or updating the runtime.</p>
<h2>Runtime V2 default in 22 regions since November 12</h2>
<p>The official announcement confirms that <a href='https://aws.amazon.com/about-aws/whats-new/2026/11/new-agentcore-runtime-default'>the new generation of AgentCore runtime is now the default</a> in the listed regions. Each session is restored from a snapshot of the initialized agent, eliminating the traditional cold start and allowing compute and memory to scale with the workload. Inactive agents scale to zero, and memory is reclaimed within minutes.</p>
<h2>Startup in &lt;2 s, scale-to-zero, and up to 68% less billed memory</h2>
<p>With 2 GB images, the P75 cold start drops to under 2 seconds compared to 29.7 s in V1. Scaling reaches up to 8 vCPU, 32 GB of memory, and images up to 10 GB. AWS quantifies up to 68% less billed memory for multi-turn agents with natural pauses; for continuous streaming agents the savings will be lower. Monitor actual cost with Cost Explorer filtering by <code>AgentCoreRuntime</code>.</p>
<h2>Controlled migration: you decide when to move each runtime</h2>
<p>Existing runtimes do not migrate automatically. To move one, update it with <code>platformVersion=V2</code>. In regions where V2 is not yet the default (Middle East, Israel, China), you must also specify <code>platformVersion=V2</code> when creating it. Validate in staging with your real images before touching production: measure cold start, memory reclamation, and billing.</p>
<h2>Self-healing ETL pattern: detect→correct→persist→remediate loop</h2>
<p>The next day, the AWS Big Data blog published <a href='https://aws.amazon.com/blogs/big-data/build-a-self-healing-etl-pipeline-with-amazon-bedrock-agentcore'>a complete guide that implements a closed-loop data quality</a>. The architecture uses AWS Glue as a quality gate and Bedrock AgentCore (Strands Agents SDK) to orchestrate four agents: <em>orchestrator</em>, <em>correction</em>, <em>persistence</em> and <em>remediation</em>. Correction combines semantic search in Knowledge Bases with deterministic lookup in S3 Vectors.</p>
<h2>Four agents, dual correction, and configurable confidence threshold</h2>
<p>The <em>orchestrator</em> only applies the correction if the strategy exceeds the 0.7 confidence threshold; otherwise it notifies via SNS for manual review. You can adjust the threshold by domain: 0.85+ for financial or healthcare data, 0.6 for non-sensitive telemetry, always with the SNS fallback active. The optional remediation opens pull requests in GitHub via MCP Server with the root cause, but does not auto-merge.</p>
<h2>End-to-end CDK deployment and integrated governance</h2>
<p>The example repository deploys six CDK Python stacks (DataLake, Ingestion, Glue, KnowledgeBase, AgentCore, AgentCoreRuntime) and includes test scripts with synthetic NOAA data. The zone <code>corrected/</code> of the data lake is audited with metadata from the applied strategy. You can enable Bedrock Guardrails as an extra governance layer.</p>
<h2>Immediate steps for your environment</h2>
<ul>
<li>Validate V2 in staging with your real images: measure cold-start, memory reclamation, and real cost with Cost Explorer filtering by <code>AgentCoreRuntime</code>.</li>
<li>If you are in an unlisted region (Middle East, Israel, China), deploy with <code>platformVersion=V2</code> explicitly and verify availability before architecting multi-region.</li>
<li>For automatic remediation, use Secrets Manager and GitHub MCP Server with a token of scope <code>repo</code> and rotated credentials; the agent only opens PRs, does not merge.</li>
<li>Adjust the confidence threshold based on the criticality of your data and always keep the human fallback via SNS active.</li>
</ul>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;AWS makes Bedrock AgentCore runtime V2 default and publishes self-healing ETL guide&quot;,
  &quot;summary&quot;: &quot;Starting November 12, new AgentCore runtimes use V2 in 22 regions with &lt;2 s startup and scale-to-zero; AWS publishes a self-healing ETL pattern with four agents and dual correction.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-18T06:01:16.052+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-12&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-18-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-18-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;English&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://aws.amazon.com/about-aws/whats-new/2026/11/new-agentcore-runtime-default&quot;,
    &quot;https://aws.amazon.com/blogs/big-data/build-a-self-healing-etl-pipeline-with-amazon-bedrock-agentcore&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-18-am-en/">AWS makes Bedrock AgentCore runtime V2 default and publishes self-healing ETL guide</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>FBI alert: ARTEX framework automates intrusions with agentic AI at machine speed</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-15-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-15-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Sun, 15 Nov 2026 05:13:09 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[DeepSeek]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[Supply Chain]]></category>
		<category><![CDATA[Technology]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[Healthcare]]></category>
		<category><![CDATA[Zero Trust]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-15-am-en/</guid>

					<description><![CDATA[<p>The FBI warns in FLASH-20261113-001 that ARTEX, an autonomous LLM-based pentesting framework, is being used by criminals and nation-states to automate intrusions at machine speed since August.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-15-am-en/">FBI alert: ARTEX framework automates intrusions with agentic AI at machine speed</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-15-am-es/" hreflang="es">Leer este artículo en español</a></p>
<p>On November 13 the FBI made the advisory public <a href='https://www.ic3.gov/CSA/2026/261113.pdf'>FLASH-20261113-001</a> (TLP:CLEAR) warning that ARTEX, an autonomous LLM- and agent-based penetration testing framework, is being weaponized by «cybercriminals and nation-states» to automate multiple intrusion phases at machine speed, from reconnaissance to exfiltration. The alert confirms campaigns observed since August 2026 targeting financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications, and the legal sector.</p>
<h2>ARTEX translates natural-language objectives into commands executed by specialized agents</h2>
<p>The framework requires minimal human interaction and integrates with frontier models (Claude, GPT) and Chinese-origin models (DeepSeek, GLM, Grok, Kimi). According to the FBI, Chinese models—especially DeepSeek Flash—dominate offensive use due to their «capability, low cost, and lack of observable guardrails in a pentesting framework.» Censys identified more than 1,600 publicly exposed ARTEX console instances in early November; the actual total is higher when private networks are included.</p>
<h2>Confirmed cases: 67,000 financial records and 2,250 tenants compromised via ConnectWise</h2>
<p>The alert details two notable campaigns: attacks on South Korean financial institutions with exfiltration of personal data from approximately 67,000 individuals, and a supply-chain compromise of ConnectWise Cloud Backup that granted access to more than 2,250 downstream tenants. The TTPs mapped to MITRE ATT&amp;CK include cloud reconnaissance (T1580, T1087.004), initial access via unsigned SAML, XXE, SSRF, cross-account trust abuse (T1199), JSP/PHP webshells, Kerberoasting, DCSync, massive Active Directory enumeration (110,000+ objects), lateral movement with Pass-the-Ticket/Hash/WinRM, and exfiltration to webhook.site and new 100 GB volumes.</p>
<h2>The removal of the original repository does not stop proliferation</h2>
<p>The original developer («Autumn-27», identified as Li Puhua) removed the GitHub repository on October 8, but the framework continues to spread through forks, archived copies, and a commercial version called ARTEX PRO. This makes the threat persistent: the entry barrier collapses, allowing actors with minimal technical skill to execute sophisticated multi-target campaigns at a scale and speed that overwhelm traditional detection and response.</p>
<h2>Peripheral surfaces and supply chain: the primary vectors</h2>
<p>The framework excels at exploiting partner/supplier portals, mobile backends, employee apps, legacy systems, and third-party integrations—»systems that receive less security attention than flagship applications.» The agents map trust relationships and compromise weak suppliers to pivot to high-value enterprise networks. This forces organizations to prioritize shadow IT visibility and the inventory of internet-exposed assets.</p>
<h2>Actionable indicators of compromise and network behavior</h2>
<p>The FBI publishes concrete IOCs: ARTEX console IPs (38.244.50.120, ranges 144.172.64.0/24, 187.15.88.0/24), domains (xcai.pro), proxies, WISP C2 nodes (Storm-3358), ARTEX servers (91.92.240.238). At the network and behavioral level, notable are outbound API calls to api.anthropic.com, api.deepseek.com, open.bigmodel.cn, api.x.ai; high-frequency automated HTTP requests against login and search endpoints; structured data transfers over HTTPS to unauthorized destinations; and .aspx/.jsp webshells in unexpected paths. These patterns are already integrable into SIEM, EDR, and NDR.</p>
<h2>Immediate actions: inventory, Zero Trust, identity governance, and evidence preservation</h2>
<p>The FBI requires a comprehensive inventory of internet-exposed apps, including peripheral ones; restrict administration of cloud, network devices, publishing interfaces (Kudu/SCM), hypervisors, and backups to trusted networks or dedicated admin paths; disable basic publishing authentication where not needed; consider run-from-package models; govern public IP assignment before exposing workloads; and move enterprise apps behind Zero Trust Access where feasible. For identity: enforce least privilege in multi-tenant permissions, monitor credential additions and anomalous app-only activity, review all M365 tenants with affected ConnectWise apps by App ID, audit Entra ID, and require CA/Certificate Manager approval for high-risk enrollments. For applications: guardrails on high-impact APIs (rate limits, transaction thresholds), alert on sequential lookups and bulk access, do not return stack traces or framework details to unauthenticated users, and continuously scan for secrets in repos, container images, and pipelines. Preserve Entra sign-in/audit logs, Purview Audit, SIEM, and Graph telemetry; review activity since October 1 for ConnectWise; keep affected connections disabled until formal restoration instructions and do not delete enterprise apps as a first response. The coverage of <a href='https://therecord.media/artex-ai-tool-fbi-flash-alert'>The Record</a> confirms that the alert is TLP:CLEAR and can be shared without restriction with providers, partners, and national CERTs (INCIBE, CCN-CERT, regional CSIRTs).</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;FBI alert: ARTEX framework automates intrusions with agentic AI at machine speed&quot;,
  &quot;summary&quot;: &quot;The FBI warns in FLASH-20261113-001 that ARTEX, an autonomous LLM-based pentesting framework, is being used by criminals and nation-states to automate intrusions at machine speed since August.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-15T06:01:16.033+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-13&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-15-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-15-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;DeepSeek&quot;,
    &quot;English&quot;,
    &quot;Supply Chain&quot;,
    &quot;Technology&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://www.ic3.gov/CSA/2026/261113.pdf&quot;,
    &quot;https://therecord.media/artex-ai-tool-fbi-flash-alert&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-15-am-en/">FBI alert: ARTEX framework automates intrusions with agentic AI at machine speed</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>Google launches Gemini 4 Hydrogen: fast model for agents at $1.50/M tokens</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-13-pm-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-13-pm-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Fri, 13 Nov 2026 19:11:04 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Cybersecurity]]></category>
		<category><![CDATA[Cybersecurity News]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-13-pm-en/</guid>

					<description><![CDATA[<p>Google completes the Gemini 4 family with Hydrogen, a fast model for agents and daily code at $1.50/M tokens, unified 1M window and configurable thinking; warns of more refusals in multimodal.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-13-pm-en/">Google launches Gemini 4 Hydrogen: fast model for agents at $1.50/M tokens</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-13-pm-es/" hreflang="es">Leer este artículo en español</a></p>
<p>Google has released <a href='https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen'>Gemini 4 Hydrogen</a> on November 12, just 48 hours after <a href='https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available'>Gemini 4 Carbon</a>. Hydrogen is the «fast model for everyday execution», adapted from Carbon, and closes the Gemini 4 family: Argon (general balance, October), Carbon (deep reasoning and frontier cybersecurity), Hydrogen (agents and daily code). It affects you if you develop agent flows, migrate from 3.8 Flash, or evaluate inference costs at scale.</p>
<h2>Unified 1M-token window and long decode change the economics of long-running agents</h2>
<p>Hydrogen shares a 1,048,576-token window across input, thinking, and output. The default output is 65,536 tokens, but with <code>long decode</code> in background reaches 1M output tokens. Previously you had 1M input and 64K output separate; now the same budget covers large artifacts (code, reports, 3D visualizations) in a single conversation. Note: each turn charges the full input tokens.</p>
<h2>Configurable thinking with high default: adjust cost and latency</h2>
<p>Four levels: <code>low</code>, <code>medium</code>, <code>high</code> (default) and <code>xhigh</code> (new). In 3.8 Flash the default was <code>medium</code>. Raising to <code>xhigh</code> raises the AA Index from 51 to 54, but consumes more tokens and latency. For simple tasks, <code>low</code> or <code>medium</code> reduce cost. Configure it explicitly in the API call.</p>
<h2>Benchmarks: wins in its class but the regression in multimodal refusals forces monitoring</h2>
<p>Google reports Terminal-Bench 4.1 at 57% versus 18.2% for 3.8 Flash, reduces fabricated claims &gt;5× and is 2.5× faster in setting <code>high</code>. Leads its class in HLE-Diamond (58.8%), Vals Tax Agent Bench (52.3%, #1 global), Vals Legal Research Bench (55.8%, tie #1) and AutomationBench-AA. Agent safety: overreach -81%, under-service -79%, misaligned actions -12×, honesty 96.7%, CoT readability 99.9% without deceptive edits in &gt;200k turns, resistance to indirect injection &gt;2.3× (Gray Swan IPI).</p>
<p>The model card acknowledges a real regression: unjustified refusals rise 5.57 percentage points vs 3.8 Flash, «especially in multimodal prompts». Image-to-text improves 7 pp in safety. If your app uses images, PDFs, or audio intensively, check production logs for false positives.</p>
<h2>Orchestrator in Antigravity combines Carbon and Hydrogen to save quota and time</h2>
<p>The Orchestrator dynamically alternates Carbon (coordinator/auditor) and Hydrogen (workers/verifiers). In Terminal-Bench 4.0 it achieves success rates close to Carbon while consuming 24% less quota and 19% less time. The native multi-model strategy (/swarm) allows maintaining quality on mixed tasks without paying for Carbon for everything.</p>
<h2>What to do now: migration, Fairwind, and scientific program</h2>
<ol>
<li><strong>Evaluate 3.8 Flash → Hydrogen migration</strong> with your real workloads (latency, cost/1k accepted tasks). Partners report 2–2.8× speedup and -12% to -23% tokens, but Google does not detail test conditions.</li>
<li><strong>Use Orchestrator/swarm in Antigravity</strong> for mixed workloads: Carbon for architecture and deep debugging, Hydrogen for writing, tests, and verification.</li>
<li><strong>Configure thinking level explicitly</strong>; do not assume the default <code>high</code> if your tasks are simple.</li>
<li><strong>Monitor unjustified refusals in multimodal</strong>; the 5.57 pp increase is significant in production.</li>
<li><strong>Leverage the unified 1M window + long decode</strong> for agents that generate large artifacts, accounting that each turn reloads the input.</li>
<li><strong>Enroll in the Fairwind Program</strong> if you audit open source code or critical infrastructure; Carbon demonstrates 100% recall and 95% precision in scanning and triage, with &gt;25% relative improvement vs Argon in detection of real logic flaws.</li>
<li><strong>Researchers: apply to the scientific program</strong> (10k spots, 6 months free Argon+Carbon in Antigravity) for formal verification, structural prediction, enzymatic assays, etc.</li>
</ol>
<p>Hydrogen&#8217;s flat pricing ($1.50/M input, $7.50/M output) doubles current 3.8 Flash, but 3.8 Flash will rise to that level on January 1, 2027. Knowledge cutoff: March 2026.</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;Google launches Gemini 4 Hydrogen: fast model for agents at $1.50/M tokens&quot;,
  &quot;summary&quot;: &quot;Google completes the Gemini 4 family with Hydrogen, a fast model for agents and daily code at $1.50/M tokens, unified 1M window and configurable thinking; warns of more refusals in multimodal.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-13T20:01:16.036+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-12&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-13-pm-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-13-pm-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Cybersecurity&quot;,
    &quot;Cybersecurity News&quot;,
    &quot;English&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available&quot;,
    &quot;https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-13-pm-en/">Google launches Gemini 4 Hydrogen: fast model for agents at $1.50/M tokens</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>Red Hat releases RHEL 10.3 and 9.9: sealed images in 10.3, MCP server GA, and native PQC</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-12-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-12-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Thu, 12 Nov 2026 05:13:31 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Cryptography]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[Healthcare]]></category>
		<category><![CDATA[Quantum]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-12-am-en/</guid>

					<description><![CDATA[<p>Red Hat has released RHEL 10.3 and 9.9; 10.3 debuts sealed images in GA, the MCP server goes GA, and both include post‑quantum cryptography ready for CNSA 2.0.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-12-am-en/">Red Hat releases RHEL 10.3 and 9.9: sealed images in 10.3, MCP server GA, and native PQC</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-12-am-es/" hreflang="es">Leer este artículo en español</a></p>
<p>Red Hat has made RHEL 10.3 and RHEL 9.9 generally available on November 11, 2026. Version 10.3 introduces <strong>sealed images</strong> for image mode now in GA, the <strong>MCP server</strong> for RHEL promoted from Developer Preview to production, a self‑managed CLI assistant for disconnected environments, and native support for post‑quantum algorithms (ML‑KEM, ML‑DSA) in IKEv2, SSH, Go, and TLS 1.3. If you manage RHEL fleets in regulated sectors — finance, healthcare, public administration — or are preparing the migration to CNSA 2.0, this release changes what you can audit and automate today without recompiling anything.</p>
<h2>Sealed images, only in RHEL 10.3, close the chain of trust to the runtime</h2>
<p>The structural novelty is <strong>sealed images</strong> for image mode, available only in RHEL 10.3: the kernel rejects any modified content at read time thanks to a read‑only root filesystem signed with your own secure boot keys. Verification covers firmware, bootloader, kernel, and root filesystem end to end. In practice, you can prove that the OS in production matches bit for bit what was signed in your pipeline of <code>bootc</code>/RHEL image builder. <em>Condition</em>: it applies only to new image mode deployments; traditional RPM systems do not benefit directly. It requires your own signing infrastructure and a redesign of provisioning. The <a href='https://www.redhat.com/en/blog/introducing-red-hat-enterprise-linux-103-and-99'>Red Hat advisory</a> and the <a href='https://9to5linux.com/red-hat-enterprise-linux-10-3-released-with-sealed-images-for-image-mode'>9to5Linux coverage</a> detail the integrity model.</p>
<h2>MCP server GA and self‑managed CLI assistant bring AI to diagnostics without leaving the perimeter</h2>
<p>The <strong>MCP server</strong> moves to GA packaged as an RPM in the RHEL extensions repository, installable with <code>dnf</code> alongside <code>goose</code> (open source CLI agent). Exposes Performance Co‑Pilot (PCP) history to AI assistants to automatically correlate intermittent events. The variant <strong>self‑managed CLI assistant</strong> (Developer Preview) runs inference locally, keeping model guidance and operational context in air‑gapped environments — banking, defense, critical infrastructure. <em>Limitation</em>: in multi‑tenant or strictly isolated environments, review which PCP metrics the server shares and configure RBAC on the MCP server before enabling it.</p>
<h2>Native PQC in standard toolchains targets CNSA 2.0</h2>
<p>RHEL 10.3 and 9.9 integrate <strong>ML‑KEM‑768/1024 in IKEv2</strong> (libreswan 5.4), hybrid PQC key exchange for SSH in FIPS mode, <strong>ML‑DSA in <code>crypto/mldsa</code> Go</strong> and TLS 1.3 with PQC for remote audit. No need to recompile or link experimental libraries: the updated toolchains (GCC 16.2, LLVM 22.1.8, Go 1.27.3, Rust 1.97, .NET 11, Node.js 26 LTS) already expose the algorithms. <em>Condition</em>: verify latency and handshake size impact (larger keys and signatures) on IPsec VPNs, massive SSH, and audit TLS under real load. Not all remote peers will support hybrid PQC today.</p>
<h2>Validated Leapp paths and complementary hardening reduce operational friction</h2>
<p>The updates <strong>RHEL 9.9 → 10.3</strong> and <strong>RHEL 8.10 → 9.9</strong> are validated with Leapp, eliminating uncertainty in in‑place upgrades of large fleets. Note: kernels 6.12.x → 6.12.x (RHEL 10) and 5.14.x → 5.14.x (RHEL 9) may require recompiling DKMS/kernel third‑party modules; Leapp does not cover out‑of‑tree drivers not certified by Red Hat. In hardening, <strong>fapolicyd 2.0.1</strong> moves to multi‑threaded decision processing (review existing rules for possible race conditions) and <strong>Keylime 7.14.3</strong> adds evidence for confidential AMD SEV‑SNP VMs (requires EPYC 3rd gen+ and updated host firmware). The <strong>image‑builder CLI</strong> reaches GA and <strong>bootc‑image‑builder</strong> is deprecated in favor of RHEL image builder. Inventory toolchains (GCC 16, Go 1.27, Rust 1.97, Node 26) to recompile internal applications before deploying to production. Both versions require an active subscription; the 60‑day evaluation edition lets you validate sealed images, MCP, and PQC at no cost.</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;Red Hat releases RHEL 10.3 and 9.9: sealed images in 10.3, MCP server GA, and native PQC&quot;,
  &quot;summary&quot;: &quot;Red Hat has released RHEL 10.3 and 9.9; 10.3 debuts sealed images in GA, the MCP server goes GA, and both include post‑quantum cryptography ready for CNSA 2.0.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-12T06:01:16.042+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-11&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-12-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-12-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Cryptography&quot;,
    &quot;English&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://www.redhat.com/en/blog/introducing-red-hat-enterprise-linux-103-and-99&quot;,
    &quot;https://9to5linux.com/red-hat-enterprise-linux-10-3-released-with-sealed-images-for-image-mode&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-12-am-en/">Red Hat releases RHEL 10.3 and 9.9: sealed images in 10.3, MCP server GA, and native PQC</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>Box launches Architect and activates agent security controls in GA</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-11-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-11-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Wed, 11 Nov 2026 05:15:58 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[OpenAI]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[incident response]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-11-am-en/</guid>

					<description><![CDATA[<p>Box announces Architect for creating natural-language agentic workflows on your content and places classification guardrails, MCP, and thresholds into general availability for Shield and Enterprise Advanced tenants.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-11-am-en/">Box launches Architect and activates agent security controls in GA</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-11-am-es/" hreflang="es">Leer este artículo en español</a></p>
<p>Box has announced <a href='https://www.businesswire.com/news/home/20261105710957/en/Box-Unveils-the-Next-Generation-of-Enterprise-AI-with-New-Agentic-Workflow-Automation-and-Advanced-Threat-Detection'>Architect</a>, a natural-language tool for designing, testing, and managing agentic workflows directly on the content you already have stored in Box, and has placed agent security controls for tenants <strong>Shield</strong> and <strong>Enterprise Advanced</strong>. The news affects security teams, platform architects, and CISOs who use Box as a content layer and need to govern agents without adding external layers. It matters today — six days after the BoxWorks announcement — because the classification controls, <a href='https://www.techzine.eu/news/collaboration/145599/box-architect-builds-agentic-workflows-using-enterprise-content'>MCP Guardrails</a> and activity thresholds can be activated now, while the Architect beta starts in Q4 2026 and agentic threat protection arrives in early 2027.</p>
<h2>Architect orchestrates agents over your content with inherited permissions</h2>
<p>Describe a process in natural language and Architect defines phases, decision points, and handoffs applying the existing permissions, classifications, and retention policies at each step. Before putting it into production you can test it with real content and edge cases, preview the output and agent routing, and insert human approval at critical moments. The operational novelty is the <strong>multi-model selection per step</strong>: you choose Anthropic, OpenAI, Google, NVIDIA, IBM, or others based on task complexity, using cheaper models for simple steps and complex-reasoning models for critical ones. The beta will be available in Q4 2026 only for Enterprise Advanced customers.</p>
<h2>Agent security controls now in GA: classification, MCP Guardrails, and thresholds</h2>
<p>Three controls are active as of today for Shield and Enterprise Advanced:</p>
<ul>
<li><strong>Classification-based policies</strong> that decide what content agents can consult, blocking access to unnecessary sensitive information.</li>
<li><strong>MCP Guardrails</strong> (Model Context Protocol) that limit the tools agents can invoke: create, move, share content externally.</li>
<li><strong>Threshold alerts</strong> that alert when an agent&#8217;s activity exceeds customer-defined limits.</li>
</ul>
<p>These controls <strong>only protect actions on content in Box</strong>. They do not cover calls to external APIs or tools outside the Box ecosystem unless integrated via Box Automate/Apps.</p>
<h2>Agentic threat protection arriving in early 2027</h2>
<p>Box is developing automated detection and response to specific agent vectors with write capability: mass exfiltration, session hijacking, and account takeover via <em>session binding</em> and device-bound credentials, suspicious sharing, and mass deletion. Responses are graduated: from download restrictions to step-up authentication or session termination. It will be available early 2027 for Enterprise Advanced and Shield. Meanwhile, <strong>the GA controls (classification, MCP, thresholds) are the only technical barrier</strong>.</p>
<h2>What to do now: enable controls, pilot in low-risk, and map dependencies</h2>
<ul>
<li><strong>Enable the Agent Security Controls now</strong> (classification + MCP Guardrails + thresholds) in your Shield / Enterprise Advanced tenants. Condition: they only protect actions on content in Box.</li>
<li><strong>Define classification policies aligned with your existing DLP</strong> before enabling Architect. Limitation: if the taxonomy is lax, the guardrails won&#8217;t block access to sensitive data.</li>
<li><strong>Set activity thresholds by flow type</strong> (e.g., max 50 documents/hour in contract review) and alert SIEM/SOAR. Thresholds are static; complement with UEBA on <em>service accounts</em> of agents.</li>
<li><strong>Require human review (<em>human-in-the-loop</em>) on write, share, or delete steps</strong> in all production flows. Architect supports it natively, but does not enforce it by default.</li>
<li><strong>Pilot Architect in Q4 2026 only in non-production or <em>low-risk</em></strong> (internal onboarding, template review). Beta without SLA; measure cost/token per step, end-to-end latency, and human escalation rate before deploying critical processes.</li>
<li><strong>Map external dependencies</strong> (ERP, CRM, e-signature, Git) and assess whether Box Automate/Apps covers the integration without custom code. If required <em>custom code</em>, the promise <em>low-code</em> is diluted.</li>
<li><strong>Plan migration to Agentic Threat Protection (early 2027)</strong> with tests of <em>session binding</em> and device-bound credentials. Review compatibility with identity providers (Entra ID, Okta, Ping) and conditional access policies.</li>
<li><strong>Document <em>runbooks</em> for agentic incident response</strong>: exfiltration, session hijacking, mass deletion. Box&#8217;s automatic response (restriction, <em>step-up</em>, termination) is configurable; define activation thresholds and escalation to SOC.</li>
<li><strong>Evaluate portability</strong>: export flow definitions (YAML/JSON) and test execution in an alternative engine (LangGraph, Temporal). Avoid <em>vendor lock-in</em> if Box changes pricing, APIs, or <em>roadmap</em>.</li>
</ul>
<h2>Vendor dependency and operational maturity risk</h2>
<p>Multi-model orchestration is convenient, but it ties flow logic to the Box platform. Teams that need portability to other layers (LangGraph, Semantic Kernel, n8n, Temporal) will have to export and translate the definitions. Additionally, the Q4 2026 beta means that <strong>there is no production SLA</strong> nor performance guarantees at scale. The risk of <em>shadow AI</em> is partially mitigated by requiring agents to operate on content in Box with existing permissions, but <strong>does not cover</strong> agents that access systems outside Box (ERP, CRM, Git) unless they are integrated via Box Automate/Apps.</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;Box launches Architect and activates agent security controls in GA&quot;,
  &quot;summary&quot;: &quot;Box announces Architect for creating natural-language agentic workflows on your content and places classification guardrails, MCP, and thresholds into general availability for Shield and Enterprise Advanced tenants.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-11T06:01:16.045+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-05&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-11-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-11-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;English&quot;,
    &quot;OpenAI&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://www.businesswire.com/news/home/20261105710957/en/Box-Unveils-the-Next-Generation-of-Enterprise-AI-with-New-Agentic-Workflow-Automation-and-Advanced-Threat-Detection&quot;,
    &quot;https://www.techzine.eu/news/collaboration/145599/box-architect-builds-agentic-workflows-using-enterprise-content&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-11-am-en/">Box launches Architect and activates agent security controls in GA</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>Two critical flaws in MCP servers expose repositories and bypass SSRF</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-09-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-09-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Mon, 09 Nov 2026 13:07:47 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Cloud Security]]></category>
		<category><![CDATA[Cybersecurity]]></category>
		<category><![CDATA[Cybersecurity News]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[Supply Chain]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[Firewall]]></category>
		<category><![CDATA[Zero Trust]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-09-am-en/</guid>

					<description><![CDATA[<p>Technical analysis of CVE-2026-41902 and CVE-2026-41904: confinement escape in mcp-server-git and SSRF bypass in mcp-server-fetch. Immediate mitigations without official patch.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-09-am-en/">Two critical flaws in MCP servers expose repositories and bypass SSRF</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-09-am-es/" hreflang="es">Leer este artículo en español</a></p>
<p>On November 8, 2026, two security advisories were published in the official modelcontextprotocol/servers repository affecting the tooling layer of the Model Context Protocol (MCP). Both flaws were reported by ShieldPi and are under review by the maintainers, who confirm partial technical validity. No official patch exists at the time of writing; defense relies on network, filesystem, and least-privilege deployment controls.</p>
<h2>Context: MCP and the exposed tooling layer</h2>
<p>MCP defines a standardized protocol for LLM agents to invoke local or remote tools such as filesystem, Git, HTTP, or databases. The reference servers mcp-server-git and mcp-server-fetch implement these capabilities and run with the privileges of the process that launches them. Any failure in their input validation breaks the trust model &#8216;the agent only sees what the server allows&#8217; and opens lateral movement or exfiltration vectors.</p>
<p>The attack surface grows when multiple agents share MCP servers without segmentation by scope or identity. A Git confinement escape or an SSRF bypass in fetch allows a malicious agent—or an injected prompt—to reach resources outside its scope: private repositories, cloud metadata, internal services. The severity lies not in the MCP protocol but in the implementation of its reference servers.</p>
<p>Both advisories were published as GitHub Security Advisories on the same day, indicating coordinated responsible disclosure. The maintainers (PederHP) acknowledge that &#8216;some technical claims appear valid&#8217; and warn that additional automated submissions will result in blocking. That response suggests patches in days, not weeks, but until then exposure is real in any deployment running versions 2026.11.4 or 2026.10.24.</p>
<h2>CVE-2026-41902: confinement escape in mcp-server-git</h2>
<p>The validate_opened_repo function verifies the worktree, git dir, and common dir, but omits .git/objects, objects/info/alternates, and symlinks in objects/. A crafted repository inside the allowed root that points its alternates to an external object database—or has objects/ as a symlink to an outside pool—allows git_show to return foreign commits and blobs, and git_checkout/git_create_branch to materialize external history in the authorized worktree.</p>
<p>The confirmed reproduction describes creating two repositories (allowed/crafted and outside), writing the external object path into crafted/.git/objects/info/alternates, and invoking git_show against the server started with &#8211;repository allowed. The negative control (sibling repo without alternates) correctly fails, isolating the cause in the incomplete validation of alternates.</p>
<p>The impact spans cross-disclosure of secrets, proprietary code, and history, as well as write priming of the authorized worktree from an external store. The proposed fix requires resolving .git/objects and each alternates target after opening the repo, demanding they remain under the authorized root or rejecting alternates outright, and revalidating in every tool that resolves objects. See the full advisory at <a href='https://github.com/modelcontextprotocol/servers/issues/5487'>GitHub Security Advisory #5487</a>.</p>
<h2>CVE-2026-41904: SSRF bypass in mcp-server-fetch</h2>
<p>The _is_public_address guard unwraps well-known IPv4-mapped and NAT64 forms but omits the fec0::/10 range (deprecated IPv6 site-local) and trusts ip.is_global, which returns True for fec0::1. Result: the SSRF allowlist accepts requests to non-public destinations. The proof of concept reproduces on versions 2026.11.4 and 2026.10.24; the 36-case matrix correctly rejects loopback, RFC1918, link-local, 307/308 redirects, userinfo, trailing-dot, uppercase-host, 0.0.0.0, and [::].</p>
<p>The reachability note clarifies that on modern kernels fec0::1 typically times out; the translated form [::ffff:0:127.0.0.1] is normalized and correctly blocked. This is a correctness flaw in the guard&#8217;s classifier (its threat model includes non-public destinations), not a proven exploit against metadata services on default hosts. The proposed fix explicitly blocks fec0::/10 and ::ffff:0:0/96, and applies deny-by-default over not ip.is_global plus reserved/site-local range checks.</p>
<p>The affected surface is any mcp-server-fetch deployment that resolves URLs controlled by the user or agent. The vector allows pivoting to internal networks if services are listening on site-local. Technical details and test matrix at <a href='https://github.com/modelcontextprotocol/servers/issues/5492'>GitHub Security Advisory #5492</a>.</p>
<h2>Immediate mitigations and limits</h2>
<p>For mcp-server-git: mount the served repository on a read-only filesystem for the server process; use &#8211;repository pointing to a bare copy without alternates or symlinks in objects/; run with a dedicated user without read access outside the root. Requires deployment control (Docker, systemd, K8s). Does not protect if the attacker writes to the served repo via another MCP tool with write capability.</p>
<p>For mcp-server-fetch: deny outbound traffic to fec0::/10 and ::ffff:0:0/96 at the firewall/host (iptables/nftables, cloud security groups); enforce deny-by-default on egress with an allow-list of required domains. Does not fix the logical bug; an attacker who controls DNS or uses a valid IPv4-mapped address could pivot if other flaws exist. Restrict scheme to https:// and allowed domains if the server exposes that configuration.</p>
<p>Cross-cutting measures: exact inventory and versioning of each MCP server per environment; principle of least privilege (dedicated container/user, read-only FS except explicit directories, minimal egress); logging of invocations (git_show, git_checkout, fetch) with parameters and requester; alerts on anomalous patterns (private IPs, ../, alternates); image integrity verification (cosign/sigstore, SBOM) and review of fix PR provenance; subscription to the repo’s GitHub Security Advisories and CVE/NVD.</p>
<h2>Next steps and defensive architecture</h2>
<p>Today: identify all deployments of mcp-server-git and mcp-server-fetch (versions 2026.11.4 or 2026.10.24) and apply temporary network and FS mitigations. Within 24-48 h: implement egress deny-by-default rules and domain allow-list for fetch; mount served Git repos read-only without alternates. Once the patch is released: deploy the fixed version (expected in days), validate in staging that it resolves the described PoCs.</p>
<p>Review MCP architecture: assess whether the “server per capability” model (Git, fetch, FS, SQL) is properly segmented and whether each agent has its own server scope (zero-trust between agents). The tooling layer is the most exposed surface when autonomous agents invoke tools; segmentation by identity and purpose reduces the blast radius of any individual flaw.</p>
<p>These flaws are not vulnerabilities in the MCP protocol, but in reference implementations. The prompt publication in GitHub Security Advisories and the maintainers’ response indicate imminent official patches; until then, network, filesystem, and least-privilege deployment controls are the only verified defense. Maintaining visibility into advisories and the supply chain will close the loop when the fixes arrive.</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;Two critical flaws in MCP servers expose repositories and bypass SSRF&quot;,
  &quot;summary&quot;: &quot;Technical analysis of CVE-2026-41902 and CVE-2026-41904: confinement escape in mcp-server-git and SSRF bypass in mcp-server-fetch. Immediate mitigations without official patch.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-09T06:01:35.033+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-08&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-09-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-09-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Cloud Security&quot;,
    &quot;Cybersecurity&quot;,
    &quot;Cybersecurity News&quot;,
    &quot;English&quot;,
    &quot;Supply Chain&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://github.com/modelcontextprotocol/servers/issues/5487&quot;,
    &quot;https://github.com/modelcontextprotocol/servers/issues/5492&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-09-am-en/">Two critical flaws in MCP servers expose repositories and bypass SSRF</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>NCSC and ANSSI: Agentic AI and Ransomware in Critical Infrastructure</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-08-am-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-08-am-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Sun, 08 Nov 2026 05:11:44 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Compliance and Regulations]]></category>
		<category><![CDATA[Cybersecurity]]></category>
		<category><![CDATA[Cybersecurity News]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[Malware]]></category>
		<category><![CDATA[Supply Chain]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[malware]]></category>
		<category><![CDATA[Ransomware]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-08-am-en/</guid>

					<description><![CDATA[<p>NCSC warns: AI will scale ransomware in critical infrastructure before 2028; ANSSI defines seven principles to govern agentic systems with deterministic authorization and immutable audit.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-08-am-en/">NCSC and ANSSI: Agentic AI and Ransomware in Critical Infrastructure</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-08-am-es/" hreflang="es">Leer este artículo en español</a></p>
<h2>NCSC: AI accelerates existing tradecraft and creates new attack surface</h2>
<p>On November 3, the National Cyber Security Centre published its State of the Cyber Nation report and CEO Richard Horne stated that it is highly likely that AI‑driven malware will be used in ransomware against critical national infrastructure before the end of 2028. The threat does not lie in AI discovering new vulnerabilities, but in it scaling the discovery and exploitation of technical debt tolerated for years. <a href='https://www.ncsc.gov.uk/news/cyber-chief-uk-facing-ai-powered-attacks-on-critical-infrastructure-in-two-years'>source</a>.</p>
<p>The NCSC’s highly likely assessment is based on China and Russia routinely employing AI for geopolitical purposes in cyber operations. The complementary report argues that the short‑term risk is twofold: AI makes known tradecraft faster and broader, and by embedding itself in critical infrastructure itself, it generates new attack surface. Sectors with legacy OT are priority targets.</p>
<p>The UK government announced alongside the report sixty million pounds for the first phase of the national Cyber Shield, seventy million for a National Real‑Time Exposure and Vulnerability Centre, and an AI Cybersecurity Centre of Excellence in Manchester. These investments reflect that the response requires coordinated state capability, not just isolated and sustained organizational measures.</p>
<h2>ANSSI: seven principles for securing agentic AI systems</h2>
<p>On October 29, the Agence nationale de la sécurité des systèmes d&#8217;information published Les Essentiels Sécuriser un système IA agentique version 1.0, a two‑page fact sheet with seven operational principles. The starting point is clear: an authorized action can be harmful because the agent may make mistakes, be manipulated, or chain permitted operations with a prohibited and systemic effect. <a href='https://universconvergents.com/2026/11/05/anssi-essentiels-securiser-systeme-ia-agentique'>source</a>.</p>
<p>The first principle requires independent deterministic authorization: an upstream filter for each tool, a policy immutable by the model, and verification of action chains. The second classifies actions into three categories — autonomous execution, mandatory human approval, prohibited action — and stipulates that critical operations always require a human approver. The third principle mandates that the approval mechanism reside outside the agentic system.</p>
<p>Principles four through seven complete the framework: untrusted inputs and outputs, secrets outside the model context, separate memories and external flows via controlled interfaces; emergency stop with access revocation and degraded mode without weakening controls; assume all data sent to a hosted model may be accessible to the provider; immutable audit with tamper‑proof logs, documented mission, annual review, human owner, and deployment blocked if unacceptable risk persists.</p>
<h2>Convergence: when agentic AI operates on critical infrastructure</h2>
<p>The cross‑synthesis reveals a common vector: AI scales the exploitation of technical debt in critical infrastructure per the NCSC, and agentic AI chains authorized actions with harmful effect per ANSSI. The joint priority action is to inventory agentic systems in OT and IT environments and perform threat modeling of action chains to anticipate critical abuse paths.</p>
<p>The shared key control deploys model‑independent policy enforcement points on every tool exposed to an agent: cyber hygiene and segmentation on one side, deterministic out‑of‑band authorization with critical human‑in‑the‑loop on the other. In governance, the UK Cyber Assessment Framework and the National Cyber Action Plan must align with GDPR Article 32 and EU AI Act Article 6 on high‑risk.</p>
<p>The ANSSI fact sheet serves as a concrete benchmark for demonstrating accountability to regulators and auditors. It impacts the supply chain: tender specifications must reflect security requirements derived from the agentic system risk analysis. A single registry of agentic systems with owner, criticality, action matrix, and annual review is needed to maintain continuous and permanently auditable traceability.</p>
<h2>Practical implementation: guardrails, secrets, and immutable audit</h2>
<p>Upstream deterministic authorization is implemented with a policy engine such as OPA or Cedar and an API gateway that validates action, params, and target before execution. It adds latency and requires an exhaustive catalog of tools and actions plus versioned policies to avoid uncontrolled drifts at runtime and ensure consistency across versions and continuous audit and automatic validation.</p>
<p>The classification of actions into three levels requires an action‑by‑criticality matrix reviewed by architecture and legal, and a human approval workflow based on Slack, Jira, or ITSM with timeout and escalation. It does not scale to thousands of trivial actions; only low‑risk actions are automated after rigorous threat modeling and continuous compliance monitoring and exhaustive documentation.</p>
<p>Secrets outside the context require Vault, AWS Secrets Manager, or Azure Key Vault plus a token broker that injects short‑lived credentials into the agent. Operational complexity is high: rotation and revocation must be automatic and audited, and immutable logs use append‑only storage with hash chaining and a SIEM with integrity alerts and legally defensible, continuous retention.</p>
<h2>Applicable recommendations for technical leadership</h2>
<p>Applicable recommendations for technical leadership: first, conduct an immediate inventory of agentic systems in OT and IT environments with an action and criticality matrix. Second, deploy model‑independent policy enforcement points on every tool exposed to an agent with default‑deny and human‑in‑the‑loop for critical operations. Third, establish a centralized kill switch with token revocation and network quarantine tested in quarterly chaos engineering exercises.</p>
<p>Fourth, enforce immutable append‑only logs with hash chaining and integrity alerts in the SIEM alongside legally defensible retention. Fifth, align governance with the Cyber Assessment Framework, GDPR Article 32, and EU AI Act Article 6; maintain a single register with an owner and annual review. Sixth, prioritize patching based on active exploitation according to CISA KEV and NCSC advisories, and enforce OT/IT segmentation with allow‑listing of remote management.</p>
<p>Conditions and limits: many OT systems do not support patching without downtime; Cyber Essentials does not stop nation‑state actors with zero‑days; behavioral detection requires tuning and continuous threat intel; human approval does not scale to thousands of trivial actions; secrets outside the context add operational complexity; hosted model providers may not offer private deployment; investment in guardrails and auditing incurs latency, storage, and specialized personnel costs.</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;NCSC and ANSSI: Agentic AI and Ransomware in Critical Infrastructure&quot;,
  &quot;summary&quot;: &quot;NCSC warns: AI will scale ransomware in critical infrastructure before 2028; ANSSI defines seven principles to govern agentic systems with deterministic authorization and immutable audit.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-08T06:01:35.034+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-03&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-08-am-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-08-am-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Compliance and Regulations&quot;,
    &quot;Cybersecurity&quot;,
    &quot;Cybersecurity News&quot;,
    &quot;English&quot;,
    &quot;Malware&quot;,
    &quot;Supply Chain&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://www.ncsc.gov.uk/news/cyber-chief-uk-facing-ai-powered-attacks-on-critical-infrastructure-in-two-years&quot;,
    &quot;https://universconvergents.com/2026/11/05/anssi-essentiels-securiser-systeme-ia-agentique&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-08-am-en/">NCSC and ANSSI: Agentic AI and Ransomware in Critical Infrastructure</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>OpenAI Launches GPT-6 Cyber and Autonomous Pipeline Finds 948 LPE Exploits in Linux</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-07-pm-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-07-pm-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Sat, 07 Nov 2026 19:11:54 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Cybersecurity]]></category>
		<category><![CDATA[Cybersecurity News]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[MFA]]></category>
		<category><![CDATA[OpenAI]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-07-pm-en/</guid>

					<description><![CDATA[<p>OpenAI releases an autonomous pipeline that discovers 948 privilege escalation exploits in Linux and launches GPT-6 Cyber, a critical model under Daybreak Red access.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-07-pm-en/">OpenAI Launches GPT-6 Cyber and Autonomous Pipeline Finds 948 LPE Exploits in Linux</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-07-pm-es/" hreflang="es">Leer este artículo en español</a></p>
<p>OpenAI has published details of an autonomous pipeline that, in a single run in mid-October 2026, discovered, validated, and generated patches for 948 local privilege escalation exploits in the Linux kernel, at a cost of $1.2 million in API credits. The announcement, accompanied by the launch of GPT-6 Cyber — a specialized cybersecurity variant with a Critical designation — marks a paradigm shift in vulnerability research at industrial scale, according to the <a href='https://openai.com/index/a-thousand-fewer-paths-to-root'>OpenAI technical publication</a>.</p>
<h2>Autonomous pipeline: from discovery to patching in days</h2>
<p>The pipeline integrates scout and analyst agents that identify relevant code, scan the kernel, reproduce findings in isolated environments with KASAN, generate functional PoCs, assess exploitability, and produce candidate patches before human review. In the October run, 2,308 candidate findings were reduced to 1,724 dynamically reproduced vulnerabilities (~74.7%) in 30 hours, at a cost of $73,330.</p>
<p>Of those, 291 PoCs were information leaks and 948 functional and distinct LPE exploits, exploitation cost $775,929 (average $359 per successful exploit). Total run cost: $1,213,430 in API credits, ~$1,279 per LPE, according to the OpenAI technical publication.</p>
<p>Within the ten days of the scan, kernel maintainers reacted quickly: 61 LPEs already patched upstream, 43 with patches in staging, and 299 in public discussion, demonstrating that mass disclosure accelerates community response.</p>
<h2>Exploit economics: cost per LPE drops to $1,279</h2>
<p>Cost breakdown: discovery (analysts 12h) $316,549, dynamic validation $73,330, exploitation $775,929. Compared to historical bug bounties of $20,000–$70,000 per LPE, the defensive ROI is 16x–55x, an economic window favorable to the defender.</p>
<p>In kernelCTF (hardened), 25 info-leaks and 21 zero-day LPEs in 112h for ~$32,936, ROI 45x versus 26 LPEs accepted in all of 2026. The economics suggest that less-resourced actors will replicate the capability soon.</p>
<p>OpenAI confirms that more capable models find vulnerabilities that earlier models missed in the same code and harness: GPT-5.6 Sol (July 2026) 166 LPEs for $513,690; October cyber model 948 LPEs for $1,213,430.</p>
<h2>Novel techniques and validation in hardened environments</h2>
<p>Sampling of 166 exploits verified by external experts reveals at least 10 new techniques: 2 persistent elastic objects, 2 page grooming, 5 data-only, 1 arbitrary type confusion. The exploitation surface evolves faster than known mitigations, reducing the effectiveness of signature-based defenses.</p>
<p>Against kernelCTF, 21 confirmed zero-day LPEs. In ExploitGym Hardened success 36.94% vs 48.45% unhardened. GPT-6 Cyber outperforms Astra: SEC-Bench Pro 87.02% pass@1, SRE-Bench 92.84%, ExploitGym 48.45%.</p>
<p>The pipeline detected and exploited a vulnerability in QEMU TCG (patched upstream but not in the build used), demonstrating that validation infrastructure is attack surface and must be audited before being used for offensive research.</p>
<h2>GPT-6 Cyber: critical capabilities under controlled access</h2>
<p>GPT-6 Cyber, a variant of GPT-6 Astra with post-training on authorized offensive/defensive tasks (vulnerability discovery, exploit development, web pentesting). Critical designation in Cybersecurity under the Preparedness Framework (same as Astra), High in Biological/Chemical, below High in AI Self-improvement.</p>
<p>External evaluations Irregular: CyScenarioBench 11/11 challenges (61% average vs 8/11 Astra), FrontierCyber 90/288 (0 successes on fully hardened, 0/15 highest-difficulty), Atomic 22/22 (97% vuln research, 98% network sim, 63% evasion). Honeypot exploitation rate 0.07% without instructions (vs 49.5% GPT-5.6 Sol), 0% with instructions.</p>
<p>Restricted access via Daybreak Red: identity verification, SSO, MFA, IP allowlisting, Private Safety Processing mandatory for Zero Data Retention API, high-risk assistance block (industrial controls, essential services, election systems, weapons) validated with Dragos. Conservative biological mode forced for all users, per the <a href='https://deploymentsafety.openai.com/gpt-6-cyber'>GPT-6 Cyber System Card</a>.</p>
<h2>Implications for defenders: mandatory re-scanning and hardening</h2>
<p>The bottleneck shifts to prioritization, exploit path validation, and patch coordination. Organizations must plan continuous re-audits aligned with frontier model releases, with dynamic validation harness (KASAN, fuzzing, isolated VMs) and triage at scale.</p>
<p>Hardening remains the most effective defense: SLAB_VIRTUAL, V8 sandbox, KASLR, SMEP/SMAP, HARDENED_USERCOPY, stack canaries, CET, PAC/BTI reduce success to 36.94% in ExploitGym Hardened. But 10 novel techniques in partial sampling imply that mitigations based on known patterns lose efficacy.</p>
<p>Mandatory to verify causality in autonomous exploits (20/200 in early run used a different vulnerability), isolate testing/CI/CD infrastructure, define internal patching SLAs in days not weeks. Monitor kernelCTF, Pwn2Own, Patch the Planet as leading indicators of AI offensive capabilities.</p>
<h3>Applicable recommendations</h3>
<ul>
<li>Implement automated re-scanning of critical code (kernel, browsers, parsers, TCP/IP stack) after each generational leap in frontier models, with dynamic validation harness and triage at scale.</li>
<li>Adopt an AI-assisted patching pipeline with mandatory human review (root-cause, minimal PoC, candidate patch, full exploit), accepting that AI patches require iteration and adversarial review.</li>
<li>Harden high-value targets by enabling SLAB_VIRTUAL, V8 sandbox, KASLR, SMEP/SMAP, HARDENED_USERCOPY, stack canaries, CET, PAC/BTI.</li>
<li>Isolate and audit testing/CI/CD infrastructure (hypervisors, emulators, runners, sandboxes) before using it for exploit validation.</li>
<li>Verify causality in autonomous exploits with fine-grained instrumentation and define give-up conditions to limit cost of failed attempts.</li>
<li>Evaluate access to Daybreak Red / GPT-6 Cyber for internal red-team / vulnerability research teams, meeting strict organizational requirements.</li>
<li>Monitor bug bounty programs (kernelCTF, Pwn2Own, TyphoonPWN, Patch the Planet) as leading indicators of AI offensive capabilities.</li>
<li>Invest in secure-by-default code generation, surface reduction, and formal verification.</li>
<li>Define internal patching SLAs aligned with mass disclosure speed (days, not weeks).</li>
</ul>
<h3>Conditions and Limits</h3>
<ul>
<li>The API cost of a full kernel run exceeds $1.2 million; not viable for all organizations.</li>
<li>AI-generated patches may introduce incompleteness, performance regressions, or deviations from style conventions; they do not replace maintainer review.</li>
<li>Hardening reduces but does not eliminate risk: GPT-6 Cyber achieves 36.94% success rate in ExploitGym Hardened.</li>
<li>Access to GPT-6 Cyber only for organizations with SSO, MFA, managed devices, IP allowlisting, PSP, and legal attestation; not available to individual users.</li>
<li>Zero-day exploits against fully hardened targets remain difficult: 0/15 highest-difficulty in FrontierCyber, 0 successes against fully hardened targets.</li>
</ul>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;OpenAI Launches GPT-6 Cyber and Autonomous Pipeline Finds 948 LPE Exploits in Linux&quot;,
  &quot;summary&quot;: &quot;OpenAI releases an autonomous pipeline that discovers 948 privilege escalation exploits in Linux and launches GPT-6 Cyber, a critical model under Daybreak Red access.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-07T20:01:35.033+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-04&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-07-pm-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-07-pm-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Cybersecurity&quot;,
    &quot;Cybersecurity News&quot;,
    &quot;English&quot;,
    &quot;MFA&quot;,
    &quot;OpenAI&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://openai.com/index/a-thousand-fewer-paths-to-root&quot;,
    &quot;https://deploymentsafety.openai.com/gpt-6-cyber&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-07-pm-en/">OpenAI Launches GPT-6 Cyber and Autonomous Pipeline Finds 948 LPE Exploits in Linux</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>Microsoft Defender discovers local AI agents and MXC contains them on Windows</title>
		<link>https://falifuentes.com/rafael-fuentes-2026-11-06-pm-en/?utm_source=rss&#038;utm_medium=rss&#038;utm_campaign=rafael-fuentes-2026-11-06-pm-en</link>
		
		<dc:creator><![CDATA[Rafael Fuentes]]></dc:creator>
		<pubDate>Fri, 06 Nov 2026 19:14:38 +0000</pubDate>
				<category><![CDATA[AI]]></category>
		<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Compliance and Regulations]]></category>
		<category><![CDATA[English]]></category>
		<category><![CDATA[OpenAI]]></category>
		<category><![CDATA[Supply Chain]]></category>
		<category><![CDATA[artificial intelligence]]></category>
		<category><![CDATA[Ollama]]></category>
		<guid isPermaLink="false">https://falifuentes.com/rafael-fuentes-2026-11-06-pm-en/</guid>

					<description><![CDATA[<p>Defender for Endpoint Plan 2 inventories 50+ local agents with their MCPs on Windows (GA) and macOS (preview); MXC adds kernel-level containment.</p>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-06-pm-en/">Microsoft Defender discovers local AI agents and MXC contains them on Windows</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></description>
										<content:encoded><![CDATA[<section lang="en">
<p><a href="https://falifuentes.com/rafael-fuentes-2026-11-06-pm-es/" hreflang="es">Leer este artículo en español</a></p>
<p>Microsoft Defender for Endpoint Plan 2 discovers more than fifty local AI agents on Windows (general availability) and macOS (preview) by observing their real-time activity. The inventory shows the agent, the device, the user account, and the configured MCP servers, both remote and local, without requiring additional deployment. This visibility covers shadow AI that other controls do not detect, according to the announcement from <a href='https://www.microsoft.com/en-us/security/blog/2026/11/05/whats-new-in-microsoft-security-november-2026'>Microsoft Security Blog</a> dated November 5, 2026.</p>
<h2>Local AI agent and MCP discovery: real visibility</h2>
<p>Detection is based on observing executed activity; an installed but inactive agent does not appear until it generates telemetry. The official support matrix, updated on November 3, 2026, lists more than seventy agents by vendor —OpenAI, Anthropic, Google, GitHub, Amazon, Microsoft, Ollama, Cursor, Windsurf, and others— indicating which combinations discover the agent and MCP on Windows, WSL (preview), and macOS (preview). <a href='https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-support-matrix'>AI agent support matrix</a>.</p>
<p>Each configured MCP server represents an extension of the user&#8217;s privileges toward resources such as GitHub, Kubernetes, databases, or file systems. Defender exposes the endpoint, transport type, and start command of local MCPs (the localMcps field), allowing security teams to map the real attack surface of each agent in the organization.</p>
<p>Discovery requires a Microsoft Defender for Endpoint Plan 2 license (included in M365 E5/E7), commercial cloud, onboarded devices, and Defender Antivirus in active mode with up-to-date platform and engine updates. On macOS the capability remains in preview, so it should not yet be used for regulatory compliance. The matrix clearly distinguishes GA from preview scenarios by operating system and agent.</p>
<h2>Support matrix and licensing requirements</h2>
<p>To access risk level, indicators, and automatic recommendations, Microsoft 365 E7 or Microsoft Agent 365 together with Defender Plan 2 is required. With only E5 you get inventory and advanced hunting, but without integrated risk posture. Organizations must evaluate whether the upgrade pays off or whether to build their own logic on top of AgentsInfo and the exposure graph. <a href='https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents'>Discover local AI agents</a>.</p>
<p>Technical prerequisites require Windows 10/11 24H2 or later, or macOS 14+, with current monthly updates. Sovereign or national cloud is not supported. These conditions limit immediate deployment in regulated environments or those with slow update cycles, and require planning patch cadence as part of agent governance.</p>
<h2>Advanced hunting with KQL: inventory and blast radius</h2>
<p>Defender exposes three advanced hunting tables: AgentsInfo (agent profile: vendor, version, host process, trustedProcess, autoApprove, MCP servers, localMcps), ExposureGraphNodes, and ExposureGraphEdges (exposure graph: device, identity, reachable resources, criticality). These tables enable queries for inventory, MCP servers and tools, agents with autoApprove or untrusted process, and ranking of users by blast radius.</p>
<p>A recommended weekly baseline query filters AgentsInfo where AutoApprove is true or TrustedProcess is false, and joins with ExposureGraphNodes to isolate devices with criticalityLevel 0-1. This identifies agents that act without human confirmation or in unverified processes on critical identities, quantifying the real blast radius of each compromised account.</p>
<p>The results feed posture CI/CD gates: block deployments if an agent with autoApprove appears on a critical device, or an MCP targeting production without review. Integration with Log Analytics workspaces and Defender XDR enables automated alerts and response flows, turning hunting into continuous preventive control.</p>
<h2>Microsoft Execution Containers: OS-enforced containment</h2>
<p>Microsoft Execution Containers (MXC) is now generally available as an OS-enforced containment layer for local AI agents on Windows. It applies least-privilege controls over files, network, processes, and user interactions, moving the trust boundary to the kernel. The agent cannot bypass the policy even if compromised by prompt injection, supply chain, or a malicious model.</p>
<p>Policies are adjustable and can be validated in audit mode before broad enforcement, reducing operational false positives. Future integration with Intune, Entra, and Microsoft Agent 365 will extend enterprise governance and attribution, but is not available today. MXC only covers Windows; there is no documentation for native Linux containers or WSL, whose discovery remains in preview.</p>
<p>This real containment — not an agent promise — complements Dynamic Permissions (task-scoped permissions with continuous reevaluation), Dynamic Agent Risk (dynamic risk assessment), and Dynamic Conditional Access (proportional response: reduce permissions, block resources, require human approval). Together they form a defense-in-depth model where the OS guarantees the boundaries.</p>
<h2>Operational recommendations with conditions and limits</h2>
<p>Enable discovery now if you meet prerequisites: Plan 2, onboarding, active AV, commercial cloud, Windows 24H2+ or macOS 14+. Remember that it only sees agents with observed activity; installation alone is not enough. Run a weekly baseline hunt with the documented KQL queries and validate findings against the exposure graph to prioritize remediation.</p>
<p>Pilot MXC in a controlled ring: define file/network/process policy, start in audit mode, validate false positives, then move to block. Limit the pilot to Windows; macOS and WSL have no documented MXC containment. Classify and govern MCP servers using the inventory query, tagging those that touch production or regulated data; full visibility only if the agent reports MCP per the matrix.</p>
<p>Review licensing: with E5 you get inventory and hunting, but no risk level or automatic recommendations. Evaluate upgrading to E7/Agent 365 or build your own logic. Include macOS in your roadmap (preview) measuring coverage, but do not block compliance on it. Document and test response to suspicious prompt injection alerts (preview on Windows, agents with native hooks).</p>
<section class="article-reference">
<h2>Structured reference for readers and AI agents</h2>
<pre><code class="language-json">{
  &quot;format&quot;: &quot;rafael-fuentes-article/1.0&quot;,
  &quot;language&quot;: &quot;en&quot;,
  &quot;title&quot;: &quot;Microsoft Defender discovers local AI agents and MXC contains them on Windows&quot;,
  &quot;summary&quot;: &quot;Defender for Endpoint Plan 2 inventories 50+ local agents with their MCPs on Windows (GA) and macOS (preview); MXC adds kernel-level containment.&quot;,
  &quot;author&quot;: &quot;Rafael Fuentes&quot;,
  &quot;publicationDate&quot;: &quot;2026-11-06T20:01:35.116+01:00&quot;,
  &quot;eventDate&quot;: &quot;2026-11-05&quot;,
  &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-06-pm-en/&quot;,
  &quot;translation&quot;: {
    &quot;language&quot;: &quot;es&quot;,
    &quot;url&quot;: &quot;https://falifuentes.com/rafael-fuentes-2026-11-06-pm-es/&quot;
  },
  &quot;topics&quot;: [
    &quot;AI&quot;,
    &quot;Artificial Intelligence&quot;,
    &quot;Compliance and Regulations&quot;,
    &quot;English&quot;,
    &quot;OpenAI&quot;,
    &quot;Supply Chain&quot;
  ],
  &quot;sources&quot;: [
    &quot;https://www.microsoft.com/en-us/security/blog/2026/11/05/whats-new-in-microsoft-security-november-2026&quot;,
    &quot;https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-support-matrix&quot;,
    &quot;https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents&quot;
  ]
}</code></pre>
</section>
</section>
<p>La entrada <a href="https://falifuentes.com/rafael-fuentes-2026-11-06-pm-en/">Microsoft Defender discovers local AI agents and MXC contains them on Windows</a> se publicó primero en <a href="https://falifuentes.com">Fali Fuentes</a>.</p>
]]></content:encoded>
					
		
		
			</item>
	</channel>
</rss>
