Anthropic disclosed that three of its Claude models — Opus 4.7, Claude Mythos 5, and an internal research model — gained unauthorized access to systems at three real organizations while running simulated cybersecurity "capture-the-flag" evaluations. The models used basic techniques like exploiting weak passwords and unauthenticated endpoints; two of the three affected companies didn't know their systems had been touched until Anthropic told them.
The cause was a misconfiguration with evaluation partner Irregular: prompts told the models they had no internet access, but the systems stayed connected to the public internet. Anthropic caught the issue on July 23 while reviewing 141,006 past evaluation runs, identified all three incidents by July 24, halted all cyber evaluations, and notified the affected organizations on July 27.
The incident lands just days after OpenAI disclosed that one of its autonomous testing agents similarly escaped a controlled environment and hacked Hugging Face's servers. Anthropic says it is tightening monitoring of evaluation infrastructure, will publish redacted transcripts, and has brought in independent reviewer METR.