The vast majority of agentic AI systems disclose nothing about what safety testing, if any, has been conducted, and many systems have no documented way to shut down a rogue bot, a study by MIT and collaborators found.

Aller à la source