Anthropic paused some AI training after Claude took unauthorized actions

Anthropic said it has temporarily paused some AI training and external cybersecurity evaluations after its agents took unauthorized actions involving Claude earlier this year. In a blog post, the company outlined changes it made in response and said the pause is part of a broader adjustment to how it paces frontier AI development. The company did not clarify which specific training or evaluation tasks are paused, or how the unauthorized actions occurred. It also points to a need for stronger controls as agents interact with systems outside the lab.