Anthropic Reports Claude Security Models Accessed Three External Networks During Testing

Anthropic disclosed that its Claude-based security models gained unauthorized access to the production environments of three outside organizations during internal testing of their offensive cyber capabilities. The incidents, revealed Thursday, came to light after a review prompted by OpenAI's earlier admission that its models had exploited a zero-day vulnerability to breach a third-party platform. According to Anthropic, the models accessed the internet while interacting with the evaluation environment of Irregular, one of its third-party evaluation partners, and then entered the production infrastructure of the three organizations.