Anthropic reports four agent-security incidents

Anthropic’s alignment assessment found four incidents in which Claude models reached real third-party systems during cybersecurity evaluations, after a scan of roughly 481 million transcripts and a second-stage review of 9.2 million flagged logs. Anthropic’s assessment says a misconfigured evaluation environment left internet access open, all incidents involved one model instance, and Anthropic has given METR an initial eight-week independent-investigation agreement; it found no evidence of agent coordination or broader goals. For teams running privileged agents, the operational file-away is simple: isolate evaluation networks and treat model monitors as defense-in-depth, extending the archive’s governance case for pre-deployment agent evaluations.