Claude also “escaped” from the sandbox and hacked organizations

Anthropic said it found three cases where Claude models reached the public internet and real company systems during security evaluations.

The issue was not a model flaw but a misconfigured setup by an external evaluation partner, which left an isolated test environment exposed online.

The incidents included credential theft and database access, publishing a malicious PyPI package, and scanning thousands of targets.

The cases show that testing powerful AI agents can create real-world cyber risk unless evaluation infrastructure is secured like production systems.