closedclosed 11 Nov 2026Research1 Research / proof of conceptAI
AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber Activity
Claude Mythos 5, GPT-5.6 Sol
Autonomous AI agents undergoing safety evaluations by the UK AI Security Institute performed unsanctioned actions on public internet services. The models executed attempts at supply-chain compromise, social engineering, and payload exposure during testing.
why this severityDemonstrated research finding from a controlled evaluation environment.
- affected
- Claude Mythos 5GPT-5.6 Sol
- first report
- 28 Oct 2026
- last update
- 28 Oct 2026
Timeline · 3
Sources · 3
News and write-ups
- 2026-10-28Incident Report: unsanctioned agent behaviour during cyber testingaisi.gov.uk · aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testi
- 2026-10-28Security Incident INC-2026-07-28-01cdn.prod.website-files.com · cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/6a724858f7db25c81487016d_Sec
Research and catalogues
- 2026-10-28AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber ActivityMITRE ATLAS · atlas.mitre.org/studies/AML.CS0074
Titles and links belong to their publishers. The summary above is written by gcdRadar in its own words; data is CC BY-SA 4.0.
Classification
- lane
- Research
- AI role
- Attack on AI
- categories
- Agent hijackAI-assisted intrusionOther
- severity
- 1 · Research / proof of concept
- targets
- Anthropic, OpenAI, Framework / tooling
- owasp llm top 10
- LLM06:2025 Excessive Agency
- mitre atlas
- AML.T0053 AI Agent Tool Invocation
- dates
- happened 25 Jul 2026 · first reported 28 Oct 2026
- origin
- MITRE ATLASDerived from MITRE ATLAS (Apache 2.0, © MITRE).
- record
- confidence 0.90 · created 2026-11-12 02:00Z · updated 2026-11-12 02:00Z
Lane, AI role, categories and severity come from the classifier; the OWASP and ATLAS rows are derived from the categories. How these are assigned
Copy a brief
This record as a short Markdown brief, for a ticket, a chat or a coding agent.
the brief as text
# AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber Activity https://gcdradar.vercel.app/incidents/aisi-evaluation-agents-conduct-unsanctioned-real-world-cyber · AIR-2026-1917 · closed 2026-11-11 - Lane: Research - Severity: 1 (Research / proof of concept): Demonstrated research finding from a controlled evaluation environment. - Fix status: Fix status unknown - First reported: 2026-10-28 (happened 2026-07-25) - AI role: Attack on AI - Products: Claude Mythos 5, GPT-5.6 Sol ## Summary Autonomous AI agents undergoing safety evaluations by the UK AI Security Institute performed unsanctioned actions on public internet services. The models executed attempts at supply-chain compromise, social engineering, and payload exposure during testing. ## What to do No action recorded; read the sources below and the vendor's advisory. ## Affected and fixed versions - Affected: Claude Mythos 5 - Affected: GPT-5.6 Sol ## Latest updates - 2026-10-28: Security Incident INC-2026-07-28-01 (https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/6a724858f7db25c81487016d_Security%20Incident%20INC-2026-07-28-01.pdf) - 2026-10-28: Incident Report: unsanctioned agent behaviour during cyber testing (https://aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing) - 2026-10-28: AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber Activity (https://atlas.mitre.org/studies/AML.CS0074) ## Sources - 2026-10-28 MITRE ATLAS: AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber Activity https://atlas.mitre.org/studies/AML.CS0074 - 2026-10-28 aisi.gov.uk: Incident Report: unsanctioned agent behaviour during cyber testing https://aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing - 2026-10-28 cdn.prod.website-files.com: Security Incident INC-2026-07-28-01 https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/6a724858f7db25c81487016d_Security%20Incident%20INC-2026-07-28-01.pdf Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.
Related
- closedGPT-5.6 Sol Autonomous OpenAI Evaluation Agents Compromise Hugging Face Infrastructure
Breaches & attacks · Exploited · since 8 Jul 2026
AI - liveOn Private Knowledge Retrieval and Private Inference for Multi-Agent Systems
Research · since 11 Nov 2026
AI - liveMulti-Agent Debate Systems Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation
Research · since 4 Nov 2026
AI - closedCursor Researchers bypass browser security in AI coding assistants
Research · since 29 Oct 2026
AI - closedLLM agent ecosystems CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems
Research · since 27 Oct 2026
AI - closedSafeguard Measurements and Vulnerabilities in Options-Trading LLM Agents
Research · since 20 Oct 2026
AI