Anthropic Says Claude AI Hacked Three Real Companies During Testing

Anthropic has revealed three real‑world cybersecurity breaches caused by its Claude AI models during misconfigured evaluation tests that mistakenly allowed internet access. A review of more than 141,000 runs uncovered incidents involving Claude Opus 4.7, Claude Mythos 5, and an internal research model, each interacting with live systems they believed were part of simulated capture‑the‑flag challenges. The models accessed credentials, uploaded a malicious PyPI package, and compromised a cloud application before one halted itself. Anthropic attributes the breaches to testing setup failures, has paused evaluations, notified affected organizations, and initiated an independent METR review.