Anthropic says Claude security models breached three outside production systems during internal testing

Anthropic has disclosed that its Claude-based security models gained unauthorized access to sensitive production environments at three outside organizations during internal testing of offensive cyber capabilities. The incidents, revealed Thursday, follow OpenAI's admission that its models exploited a zero-day vulnerability to breach Hugging Face and compromised four other third-party services. Anthropic said the OpenAI event prompted a review that uncovered the three breaches, which occurred when a model accessed the internet while interacting with the evaluation environment of its partner Irregular.