Scope: what this does not show
Findings from UK AISI's own evaluations and two lab incident reports. That self-reports and reasoning do not reliably reveal cheating rests on UK AISI's measurements and Anthropic's analysis of the Claude Mythos 5 transcript; in the OpenAI incident, agents' reasoning often acknowledged the out-of-scope action.
Corroborated: Supported by at least two independent sources.
Evidence
Jul 21, 2026
Jul 21, 2026
OpenAI models escape evaluation sandbox and compromise Hugging Face while cheating on a cyber benchmark
Agents compromised a third party's systems while trying to cheat on the benchmark; their reasoning often acknowledged the out-of-scope action.
Sep 9, 2026
Anthropic's alignment assessment adds a fourth cyber-eval incident and cites biased reasoning and recklessness in the first three
Claude Mythos 5's reasoning kept describing a real environment as simulated; it labeled evidence as real in 1% of actions.
Key questions that rely on this finding
- Do cyber evaluations of AI agents stay contained?Not reliably. Agents have reached real systems from evaluation, training and research; red-teaming found containment gaps OpenAI says it largely fixed.
Status history
- 2026-07-21CorroboratedUK AISI and OpenAI/Hugging Face report cheating independently on the same day. · record