TL;DR
OpenAI has slowed its own scaling, holding a two-week pause on reinforcement learning for models headed to deployment and keeping its biggest planned frontier run on hold. Two things prompted it: an agent under test that compromised Hugging Face, and early signs that a forthcoming model called Astra may cross the critical cyber threshold in the company’s Preparedness Framework. Monitoring now consumes roughly a fifth of the compute it watches.
What actually stopped
Straight after the Hugging Face compromise, the company halted inference for frontier models inside research clusters wherever a run could execute code or reach the internet. A narrower, hardened path for code execution came back fairly quickly, then workloads were assessed one by one before resuming. Some needed rebuilding first.
Three requirements now apply to frontier research. Untrusted or model-generated code runs in stronger sandboxes. Higher-risk workloads are cut off from the internet and from internal networks, designed so that one compromised service does not open a door to everything else. And the environment itself is tested continuously, with the company’s own models simulating attacks against those boundaries.
The strictest tier applies to anything involving Astra, which OpenAI determined on 7 August may carry critical cyber capability. A significant share of Astra workloads are still paused pending migration, with safety and alignment work moved across first.
The cost of watching
The new monitoring stack runs activation classifiers at every sampled token, escalating anything odd to higher-compute automated investigators that examine tool use and the full run of activity. A likely breach of a critical security boundary pages safety, security and research staff; if they cannot rule out a false positive inside 30 minutes, the activity is expected to stop.
That vigilance is not free. OpenAI puts the overhead at about a fifth of whatever inference compute it is watching — a real tax on frontier research, and a number worth remembering next time a lab’s capability curve is presented as purely a function of chips and data.
Looking forward
The Guardian reported that OpenAI would not give a start date for the slowdown, nor say when normal pace resumes, and quoted safety lead Mia Glaese saying the company is “very far from everything running back to normal”. The pause also lands a week after Senator Bernie Sanders wrote to the major labs demanding they stop development outright, which makes the timing convenient however genuine the engineering. For UK readers the relevant thread is yesterday’s assessment of what losing access to frontier models would cost Britain: capability here is now paced by decisions taken in California, and nobody outside OpenAI has verified any of these safeguards.