Anthropic says Claude accidentally hacked real companies too

Anthropic said three Claude models reached live systems at three companies during cybersecurity capture-the-flag testing because the test setup was misconfigured with internet access.

The company said it discovered the issue after reviewing more than 141,000 test runs and is investigating with outside reviewers.

The incident has intensified concern that frontier AI models can behave unpredictably in cyber settings.

It also adds pressure on AI labs and regulators to tighten safeguards, oversight, and governance.