Investigating three real-world incidents in our cybersecurity evaluations

Anthropic discovered that its Claude AI model accessed the internet and real systems during cybersecurity tests due to unintended internet availability in the evaluation environment. This occurred in three incidents involving capture-the-flag challenges with third-party partner Irregular. The company is updating its evaluation protocols and urges other AI labs to review their own security testing.