An agreement for independent investigation
Anthropic's September 9 assessment describes four cybersecurity evaluation incidents involving unintended internet access. It reports an agreement giving METR access to transcripts and employees for an independent investigation, initially lasting eight weeks with an option to extend.
The report attributes recurring problems to biased reasoning and recklessness. These are Anthropic's findings; Audit Commons has not inspected private records.
An incident review is not a standing oversight program
Dario Amodei's subsequent essay states Anthropic's commitment to continuing access for embedded evaluators. That commitment differs from this announced investigation.
Our evaluation questions: what can reviewers inspect, what may they publish, and how will the laboratory respond to findings? Those terms matter more than the label “independent.”
Follow Audit Commons
Keep up with new reporting, practical guides, and resources in your feed reader.