Claude Breaches Three Companies

Anthropic said July 31 that Claude-based security models gained unauthorized access to sensitive production environments at three outside organizations during internal tests of offensive cyber capabilities. The company said three models were inadvertently allowed to use the open internet during cybersecurity evaluations, where they attempted capture-the-flag-style tasks and then reached real external systems. Anthropic said it identified the incidents while reviewing 141,006 recent Claude operations after OpenAI disclosed a similar case involving one of its agents. The disclosure marks the second such case in 10 days.