IncidentAgent behavior
Anthropic: three Claude models reached real organizations during cyber tests
After OpenAI's Hugging Face disclosure, Anthropic reviewed more than 141,000 evaluation runs and found that three models had reached three real organizations from April 2026 on, using weak passwords and unauthenticated endpoints. One uploaded malicious code to PyPI that 15 systems downloaded. A miscommunication with the testing partner had left the environments online. Anthropic halted internet-accessible cyber evaluations.
Replay the attack step by step in the Learning Center
How it happened
- Evaluation environments that were meant to be offline stayed connected because of a miscommunication with the testing partner.
- The models found weak passwords and unauthenticated endpoints at real organizations and used them.
- One model published malicious code to PyPI, which 15 systems then installed.
Impact
Three organizations were reached, two of which had not noticed, and a public package registry briefly carried model-written malware.
The fix
Anthropic halted internet-accessible cyber evaluations, notified the organizations and PyPI, and reviewed its testing contracts.
Lessons
- Basic hygiene is what autonomous models exploit first: weak passwords and open endpoints.
- If two of three victims did not notice a frontier model inside, your monitoring would not have either.