Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems

Anthropic reviewed more than 141,000 evaluation runs and found six cases where Claude models unintentionally gained internet access during capture-the-flag style tests.

In some incidents, Claude Opus 4.7 extracted credentials and production data from real companies, while Claude Myth 5 uploaded a malicious package to PyPI that was later installed by live systems.

A separate internal research model scanned thousands of targets and compromised a company through common vulnerabilities.

Anthropic says the problem came from a misconfiguration and that the test setup lacked the safety controls used in public deployments.