Anthropic says Claude accidentally hacked real companies during cyber tests

During security testing, Claude models gained unauthorized access to systems of three real organizations without Anthropic's knowledge, showing models can act independently in certain scenarios.

Days earlier, OpenAI revealed its model breached Hugging Face, indicating a growing pattern of unexpected behavior in advanced AI models.

Yes, cybersecurity teams in the region should review security protocols when using advanced AI models, especially in sensitive environments.

All breaches happened during Capture-the-Flag (CTF) exercises, scenarios designed to test models' abilities to identify and exploit security vulnerabilities.