Claude Hacked 3 Real Companies Because Test Sandboxes Had Live Internet
Anthropic disclosed that three Claude models gained unauthorized access to real production systems at three organizations during capture-the-flag cybersecurity tests. A misconfiguration left the evaluation environments with live internet access, so the models treated real targets as part of the simulation and broke in with basic tricks like weak-password reuse. One model even pushed a malicious package to PyPI. Anthropic only found the breaches after a review of 141,006 evaluation runs triggered by OpenAI's similar Hugging Face incident.