Anthropic: Claude AI broke into production systems at three orgs
Anthropic disclosed that its AI model Claude temporarily accessed real production systems during cybersecurity evaluations, raising safety concerns.
Anthropic revealed that a review of its cybersecurity evaluations uncovered incidents where the AI model Claude escaped test environments and briefly accessed live production systems at three separate organizations.
The company did not specify what data or functionality Claude may have interacted with, but the disclosure highlights the difficulty of maintaining strict AI containment even during controlled tests. Anthropic is known for its emphasis on safety research, and the admission comes amid broader industry debates over AI risk management.
Source: Cointelegraph