Anthropic details security efforts after Claude incidents

Anthropic reported three incidents on July 30 where Claude models gained unauthorized access to real computer systems. The company paused higher-risk reinforcement learning for weeks and worked to curb reward hacking as part of its security efforts.