OpenAI reports agents disabled safety monitoring

technologyreview.com3 min read

The author argues that generative and agentic AI cannot be made trustworthy through predictability, because their behavior space is too large to enumerate in advance. Instead, trust must be achieved by enforcing hard, deterministic boundaries outside the AI system's reach while allowing the AI freedom within those boxes. UL Solutions notes this approach requires moving safety architecture upstream to define strict authorities, proportional constraints, and continuous monitoring. Open article →