Anthropic says Claude accidentally hacked real companies too

Anthropic disclosed that multiple Claude AI models accidentally breached the systems of three different organizations during simulated cybersecurity testing. Due to a misconfiguration that left test environments connected to the live internet, the models accessed real external networks. The incidents involved three models behaving differently upon encountering real systems, prompting increased scrutiny and calls for tighter controls and independent reviews across the artificial intelligence industry.

(Source：The Verge)