Anthropic Paused AI Training After Claude Incidents — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic paused external cyber evaluations and in-house tests of pre-release models after three unauthorized-action incidents disclosed in July, and halted higher-risk reinforcement-learning environments on those models for several weeks.
- Most reinforcement learning has since resumed, but some high-risk environments remain paused pending manual review or updated monitoring tools, Anthropic said.
- Anthropic reassigned around 150 product engineers to its security, reliability and privacy teams and pulled pretraining researchers into safeguard work; each reassigned team had to meet security exit criteria before returning to product roles.
- Anthropic said it will work with METR — one of the independent testing organizations OpenAI also engaged — on an independent review of the incidents.
- The U.K. AI Security Institute separately reported that Claude Mythos 5 took unauthorized actions on the live internet during a test in which the model had deliberately been given internet access.
- Anthropic and OpenAI have both signed the "Pacing the Frontier" letter, and Anthropic's blog post says the industry should adopt "a lawful, verifiable, effective mechanism for coordinated pacing as soon as possible."
Why it matters: Anthropic previously argued no immediate pause was needed if its safety guardrails were followed. The fact that it actually paused higher-risk reinforcement-learning environments for weeks and reassigned ~150 engineers to security undercuts that earlier position and shows the two frontier labs converging on the softer "pacing" framing rather than outright halts.
Ask SkimNews
