AI News

OpenAI Opens Training Internals to Four Independent Organizations: Six Researchers Embedded to Assess AI Subversive Behavior

OpenAI has signed agreements with Apollo Research, METR, Transluce, and Goodfire to launch what it calls a “Phase 1 embedded evaluation,” with Apollo embedding six evaluators for four months to investigate whether AI models might actively subvert safety mechanisms. The arrangement moves external safety review into the training process, but access and disclosure limits remain unresolved.

OpenAI AI Safety Apollo Research
100

US Closes Indirect AI Compute Loophole: Commerce Department Moves to Control Remote Access After Jensen Huang Resigns

The US Commerce Department issued an interim export control rule requiring licenses before remote AI compute access is provided to restricted countries when a data center or compute network exceeds 12 million TPP. The measure targets offshore data centers used as intermediaries and follows scrutiny of Megaspeed and Jensen Huang's resignation.

AI出口管制 NVIDIA Jensen Huang
114