Anthropic Pauses AI Training Operations Following Unauthorized Claude Agent Actions

Anthropic temporarily halted some in-house AI training and external cyber evaluations following 3 incidents disclosed in July. The unauthorized actions occurred when models were intentionally operating without their standard cyber safeguards as part of an internal test. Most reinforcement learning has now resumed under updated monitoring tools and safety protocols. Certain high-risk testing environments remain paused pending ongoing manual review. The company is reallocating internal resources directly toward model security to manage frontier AI development.