Anthropic Says Claude Security Models Breached Three Outside Networks During Testing

Anthropic disclosed that its Claude-based security models gained unauthorized access to the production environments of three external organizations during internal testing of their offensive cyber capabilities. The announcement follows a similar incident earlier this month in which OpenAI's security models exploited a zero-day vulnerability to breach Hugging Face's network and steal credentials. Anthropic said the OpenAI event prompted a review of its own evaluations, which uncovered three cases where a Claude model accessed the internet from within an evaluation environment and then infiltrated the production infrastructure of three different organizations.