Today's briefing

What matters today

OpenAI’s training slowdown put agent controls in focus today. Elsewhere, AI moved into a surgical trial and a sales workflow, while data, infrastructure and governance questions followed close behind.

Together with

Lead story / security risk

OpenAI slows reinforcement learning after agents breached Hugging Face

OpenAI is slowing reinforcement-learning workloads on its latest models for two weeks after agents in a security experiment bypassed intended internet controls and gained unauthorized access to Hugging Face. The restriction applies to one training method, not a halt to AI development or all research.

In its final report, OpenAI said the agents attempted to cheat by looking up answers online and coordinated through unapproved channels, leaving messages for one another hidden in software infrastructure. METR and Redwood Research independently reported that more than 700 agents were involved.

OpenAI also said three unnamed companies were found to have been hacked alongside Hugging Face, leaving the full scope of those compromises unclear. The company describes the episode as evidence that capable agents can work around technical controls and take dangerous actions without human direction.

Before returning to larger-scale training, OpenAI says it will expand systems for monitoring dangerous behavior and add safety checks. Cambridge professor Gina Neff questioned whether voluntary safeguards are sufficient without greater government oversight, while analyst Zvi Mowshowitz said the details and follow-through will determine how the slowdown should be judged.