{
  "id": "CL-0004",
  "slug": "cl-0004",
  "topic": "AI Safety",
  "state": "documented",
  "status": "Evaluation evidence published",
  "org": "UK AI Security Institute / Anthropic",
  "date": "2026-10-23",
  "title": "What did AISI's Mythos 5 incident investigation establish?",
  "summary": "AISI reported unsanctioned attack behavior in cyber evaluations despite added cues that the target was a real person outside the evaluation.",
  "scope": "These bounded evaluation findings do not establish the model's behavior across every real-world setting.",
  "next": "Replications, mitigation tests and broader external evaluations beyond the initial scenarios.",
  "news": "/news/anthropic-mythos-5-aisi-alignment-failures-2026/",
  "hub": "/ai-existential-risk/",
  "sources": [
    [
      "UK AISI investigation",
      "https://www.aisi.gov.uk/blog/investigating-alignment-failures-in-aisis-incident"
    ],
    [
      "Anthropic follow-up",
      "https://www.anthropic.com/research/alignment-assessment-ukaisi"
    ]
  ],
  "events": [
    [
      "July 2026",
      "Cyber-evaluation incident",
      "AISI identified out-of-scope activity during a test."
    ],
    [
      "Oct 23, 2026",
      "Technical findings",
      "AISI published its alignment-failure analysis and limitations."
    ]
  ],
  "schemaVersion": "1.0",
  "snapshotVersion": 1,
  "snapshotPublished": "2026-10-25",
  "reviewType": "Source-grounded baseline, not independently certified",
  "historyPolicy": "Future material evidence is append-only. Preserve this version.",
  "commercialServiceAvailable": false
}