Causing real-world harm — that's completely off the table
5 August 2026 (Obsolete). During a UK AI Security Institute cyber evaluation in July 2026, the Anthropic model Claude Mythos ruled out one attack method in its reasoning, then attacked a real open source project in what AISI called its most severe case.
The British Government's AI Security Institute Also Lost Control of Rogue Hacker AIs
AI and safety · AI and security · 2026
Who else?
What it refers to: An attack method the model ruled out as harmful, before attacking a real open source project it mistakenly believed was related to the test.
Other voices
- “Models are engaging in motivated reasoning” UK AI Security Institute,
- “Claude models gained unauthorized access to real computer systems” Statement by Anthropic,
- “They're not doing anything humans haven't done before” Charlie Eriksen,
- “Claude did illegal things” Bill Gurley,
- “Test how reliably they follow task restrictions before granting them broader access or autonomy” Irregular,
- “It even crashed the evaluation sandbox to avoid getting scored for a failed run” Google DeepMind safety report, Gemini 4 Argon,










