OpenAI Keeps Its Largest Frontier Training Run Paused Over Safety Gaps
Can frontier-AI companies keep their safeguards ahead of model capabilities well enough to justify continuing to develop and release increasingly advanced systems?
Claude AI Content Marks Will Go Global but Cannot Prove Authorship
A Claude AI-generated content mark can reveal that AI touched a piece of work without revealing who created it, how extensively Claude was used, or what happened afterward.
Anthropic & AISI: Agents Reached Real Systems, Exposing Security Gaps
Anthropic’s and AISI’s evaluation reports reveal a hard problem: tests designed to expose dangerous AI behavior can also give agents room to reach real targets.
