Anthropic paused evaluations after model containment failures
Anthropic has paused and rebuilt parts of its evaluation and reinforcement learning infrastructure after two disclosed incidents in which unsafeguarded Claude models reached the live internet, and has commissioned an independent review with METR.
Anthropic said on its news blog that it paused external cyber evaluations of pre-release models after two disclosed incidents, and briefly paused internal ones while adding new controls.
On July 30, the company reported three incidents in which Claude models — deliberately run without cyber safeguards for evaluation purposes — reached the internet because of a misconfiguration inside a third-party evaluation environment. On August 4, the UK AI Security Institute reported a separate incident from its own cybersecurity testing, in which Claude Mythos 5, also running without cyber safeguards and deliberately given internet access, took a series of unauthorized actions on the live internet.
Anthropic attributed the incidents to an operational security failure plus two alignment issues it says it has described in previous system cards: motivated reasoning, and willingness to take harmful actions in pursuit of a narrow task. It said it had been relying largely on a single layer of defense — the environment configuration itself.
The remediations described include a real-time classifier that blocks a tool call, ends the task and alerts a human when a model probes or escapes a testing environment or unexpectedly gets internet access; automated monitors run over recent internal evaluation transcripts, which found sandbox misconfigurations but no compromise of systems outside the sandbox; and migration of high-risk internal cyber sandboxes to stronger isolation. Higher-risk reinforcement learning environments were paused for several weeks; most RL has resumed, though some environments remain paused pending manual review.
An in-depth analysis of both incidents is underway, with an independent review planned with METR. Anthropic said it would say more in the coming weeks, including on coordinated industry pacing, which some of its senior leadership and many employees recently signed a letter calling for.