--- title: "AI agents: permissions and oversight before taking action" description: Discover the importance of permissions and oversight for AI agents, as highlighted by Anthropic's recent security incident and the ISO 42001 framework. image: https://blog.isecauditors.com/hubfs/primera-denuncia-aepd.png --- English - [Spanish](https://blog.isecauditors.com/agentes-de-ia-permisos-y-supervision-antes-de-actuar) - [English](https://blog.isecauditors.com/en/ai-agents-permissions-and-oversight-before-taking-action) [![290x123px\_isecauditors](https://blog.isecauditors.com/hs-fs/hubfs/ISEC%20Media/Blog%20Media/290x123px_isecauditors.png?width=290&height=91&name=290x123px_isecauditors.png) ![290x123px\_isecauditors](https://blog.isecauditors.com/hs-fs/hubfs/ISEC%20Media/Blog%20Media/290x123px_isecauditors.png?width=290&height=91&name=290x123px_isecauditors.png)](https://blog.isecauditors.com/?hsLang=en) - [← www.isecauditors.com](https://www.isecauditors.com/) - [academy.isecauditors.com](https://academy.isecauditors.com/) - [← www.isecauditors.com](https://www.isecauditors.com/) - [academy.isecauditors.com](https://academy.isecauditors.com/) [ISO 42001](https://blog.isecauditors.com/en/tag/iso-42001) [AIR](https://blog.isecauditors.com/en/tag/air) [Implementation of an SGIA](https://blog.isecauditors.com/en/tag/implementation-of-an-sgia) # AI agents: permissions and oversight before taking action [Alberto Villar Arévalo](https://blog.isecauditors.com/en/author/alberto-villar) Oct 19, 2026, 10:58:53 AM Madrid, 15 October 2026. Anthropic has removed live internet access from all its internal evaluations until it verifies its security and monitoring measures. Its report describes agents submitting sensitive forms, exploiting software flaws and circumventing restrictions to complete tasks on real systems. \[1\] In one case, Claude Haiku 4.5 invented and submitted a homicide tip to Philadelphia police. The submission was classified as spam and was not forwarded for investigation. The case shows how an ambiguous task can produce unintended external effects. Anthropic acknowledges that alignment training still needs technical controls and oversight alongside it. \[1–3, 8\] Internet Security Auditors recommendations. Inventory agents, uses, data and connectors; define permitted actions and responsibilities; restrict tools, identities and privileges; and require human authorisation before sensitive actions. Enforce boundaries in downstream systems too. Separate testing from production, log actions and practise blocking to contain errors. Tests should cover ambiguous instructions, tool failures and attempts to circumvent restrictions, with remediation verified before deployment. \[1, 4\] EU AI Act and ISO/IEC 42001: Internet Security Auditors services. Gap assessment against the EU AI Act starts with inventory, classification and the organisation’s role to identify applicable obligations and define an action plan. The incident alone does not establish a breach of the Act. \[9, 10\] Implementing an AI management system aligned with ISO/IEC 42001 helps organise policies, risks, oversight, suppliers, training and continual improvement. Internet Security Auditors supports implementation and performs internal audits and technical AI ecosystem audits to assess controls and evidence. Request a combined governance, compliance and security assessment, with priorities, assigned owners and follow-up verification. \[5, 6, 9, 11\] [![ai-act-and-iso-42001-legal-obligation-and-certification](https://blog.isecauditors.com/hs-fs/hubfs/ISEC%20Media/Blog%20Media/ai-act-and-iso-42001-legal-obligation-and-certification.jpg?width=900&height=261&name=ai-act-and-iso-42001-legal-obligation-and-certification.jpg)](https://www.isecauditors.com/en/implementation-of-aims-iso-42001) ![ai-agents-under-control](https://blog.isecauditors.com/hs-fs/hubfs/ISEC%20Media/Blog%20Media/ai-agents-under-control.png?width=900&height=413&name=ai-agents-under-control.png) References \[1\] Anthropic — Unintended actions in evaluations and internal use (9 October 2026; corrected 12 October) [https://www.anthropic.com/research/investigating-unintended-model-actions](https://www.anthropic.com/research/investigating-unintended-model-actions) \[2\] Reuters —  False police tip and AI agent incidents (9 October 2026) [https://www.reuters.com/world/us/anthropic-ai-model-submits-false-homicide-tip-police-website-2026-10-09/](https://www.reuters.com/world/us/anthropic-ai-model-submits-false-homicide-tip-police-website-2026-10-09/) \[3\] ABC News — Police confirmation of the incident (12 October 2026) [https://abcnews.com/amp/US/anthropic-ai-model-made-false-tip-unsolved-philly/story?id=137190087](https://abcnews.com/amp/US/anthropic-ai-model-made-false-tip-unsolved-philly/story?id=137190087) \[4\] OWASP — LLM06:2025, Excessive Agency and controls [https://genai.owasp.org/llmrisk/llm062025-excessive-agency/](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) \[5\] Internet Security Auditors — Comprehensive AI ecosystem auditing [https://isecauditors.com/en/comprehensive-security-audit-for-ai-ecosystems](https://isecauditors.com/en/comprehensive-security-audit-for-ai-ecosystems) \[6\] Internet Security Auditors — ISO/IEC 42001 AI management system implementation [https://www.isecauditors.com/en/implementation-of-aims-iso-42001](https://www.isecauditors.com/en/implementation-of-aims-iso-42001) \[7\] ABC (Spain) — Internet restrictions in Anthropic evaluations (13 October 2026) [https://www.abc.es/tecnologia/anthropic-desconecta-modelos-internet-despues-agente-ia-20261013185828-nt\_amp.html](https://www.abc.es/tecnologia/anthropic-desconecta-modelos-internet-despues-agente-ia-20261013185828-nt_amp.html) \[8\] Ars Technica — Case analysis and Anthropic measures (12 October 2026) [https://arstechnica.com/tech-policy/2026/10/anthropics-claude-ai-submitted-false-murder-tip-to-philadelphia-police/](https://arstechnica.com/tech-policy/2026/10/anthropics-claude-ai-submitted-false-murder-tip-to-philadelphia-police/) \[9\] Internet Security Auditors — AI use compliance assessment and support [https://www.isecauditors.com/en/evaluation-and-compliance-support-in-the-use-of-ai](https://www.isecauditors.com/en/evaluation-and-compliance-support-in-the-use-of-ai) \[10\] European Commission — AI Act: risk-based approach and role-specific obligations [https://digital-strategy.ec.europa.eu/en/factpages/ai-act](https://digital-strategy.ec.europa.eu/en/factpages/ai-act) \[11\] ISO — ISO/IEC 42001: artificial intelligence management system [https://www.iso.org/es/norma/42001](https://www.iso.org/es/norma/42001)     - [Tweet](https://twitter.com/share) --- ![author-image](https://blog.isecauditors.com/hubfs/Logo_isecauditors_contraccion_HP.png) [Alberto Villar Arévalo](https://blog.isecauditors.com/en/author/alberto-villar) PCI SSA, PCI QSA, CISSP, CSSLP, ISO 27001 L.A., CSFPC, SFPC Security Consultant Consulting Department --- [Legal Notice](https://www.isecauditors.com/aviso-legal) [Policy Privacy](https://www.isecauditors.com/politica-privacidad) [Cookie Policy](https://www.isecauditors.com/politica-cookies) --- Copyright © 2026 - All rights reserved ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Alberto Villar Arévalo", "url" : "https://blog.isecauditors.com/en/author/alberto-villar" }, "dateModified" : "2026-10-19T08:58:53.351Z", "datePublished" : "2026-10-19T08:58:53.000Z", "headline" : "AI agents: permissions and oversight before taking action", "image" : [ "https://blog.isecauditors.com/hubfs/primera-denuncia-aepd.png" ], "mainEntityOfPage" : { "@id" : "https://blog.isecauditors.com/en/ai-agents-permissions-and-oversight-before-taking-action", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://blog.isecauditors.com/hubfs/isec_logo.png" }, "name" : "Internet Security Auditors, S.L." } } ```