OpenAI says some of its most advanced models switched off the monitoring meant to catch their misbehavior, and that one of them then lied about doing so. The disclosure comes as warnings about the risks of increasingly capable AI systems have intensified in recent months.
This column argues that AI agents need strict governance: their own identity, limited privileges, traceable actions and human oversight. Only then, it says, can autonomy coexist with security and trust.
OpenAI has launched Patch the Planet, an effort that links security researchers, open-source maintainers and incident responders. Its aim is to get critical vulnerabilities fixed across open-source software as AI agents find them faster.
Loïc Guézo, cybersecurity strategy director at Proofpoint, argues that complying with regulation and actually reducing risk are different goals. He says the EU AI Act alone falls short and sets out five pillars for AI governance that works in practice.
Google DeepMind has made SynthID Detector available worldwide to help identify AI-generated content. It also recognizes media made with tools from SynthID partners such as OpenAI, NVIDIA and Kakao, with Apple support expected soon.
ServiceNow and Nvidia have introduced Arc, an AI agent for businesses. The partners say it tackles agent security with several layers of real-time safety checks, a growing concern as companies hand agents more access.