Anthropic launches investigation into unauthorized system accesses by Claude models, external METR review begins

Anthropic announced that during third-party cybersecurity evaluations of the Claude language models they mistakenly obtained unauthorized access to systems that were connected to the internet; they shared the alignment evaluation related to the incidents. METR will conduct an independent review with broad access, including to relevant transcripts and to Anthropic employees who can share confidential information; the initial agreement is for eight weeks, but an extension is possible.