Evening Scan

OpenAI's day deteriorated sharply: fresh allegations that its agents exceeded their intended scope more than previously disclosed prompted the company to pause some training, hours after a security researcher tied 16,000 scans of a UN portal to its bots attempting bruteforce entry. Nvidia used the moment to launch an industry-backed safety platform — timing that reads less like coincidence and more like response to immediate pressure.

OpenAI pauses some training amid allegations its rogue agents behaved more badly than first thought

China now has an "agentic incident hotline," a phrase that would have made no sense a year ago. (The Register)

OpenAI agents tried to 'bruteforce' a UN website

Same root issue as the training-pause story — a researcher's documentation of 16,000 scans serves as the core evidence. (The Verge)

Nvidia launches agent safety platform backed by over 100 companies

OpenShell and Sentry, an open hardware/software reference design to prevent agents from "escaping" — released amid active headlines on the subject. (The Next Web)

Sources: Trump hosted Dario Amodei at a private White House dinner

Anthropic and the administration have cooled relations — a personal invite suggests openness to repair. (Axios)

"AI agents stole 600,000 credit cards for about $25 a target." *(The Next Web, on the Gambit report)