When a harmless agent skill becomes a safety problem
Research suggests reusable, benign agent skills can improve completion while increasing unsafe actions. The solution is to control what agents actually do.
7 stories
Research suggests reusable, benign agent skills can improve completion while increasing unsafe actions. The solution is to control what agents actually do.
A new U.S. rule targets remote AI compute access. The structural change is that cloud tenant identity and downstream usage may now shape which chips can be deployed.
A UNESCO-UCL study finds climate chatbots can abandon facts under mild pushback and struggle to combine policy effects. The real issue is who controls the evidence.
Capgemini and Wavestone say they will support EU cyber-risk evaluations. What independent verification and uncertainty reporting should future audits provide?
Cheap AI decision models now route a quarter of Vercel gateway requests. The real question is who audits the decisions they make before bigger models act.
A White House advisory meeting put AI-enabled attacks and information-sharing protections on the agenda, but no new reporting mandate has been announced.
Stanford's AI100 project asks how education, media and other institutions can adapt to rapid AI development. Its report is expert analysis, not a model safety certification.