OpenAI has now put its own name to a decision it made weeks earlier: hold back reinforcement learning on its newest models and leave its biggest frontier training run parked rather than let it proceed on schedule. The company laid out the pause in a post on its own site this month, tying it to two events, a security lapse tied to Hugging Face and early evidence that an unreleased model called Astra may land in the Critical tier for cyberattack capability under OpenAI’s Preparedness Framework, the company’s internal system for tracking when a model’s abilities cross specific risk thresholds.
AI Insiders reported on August 8, citing The Decoder and WIRED, that OpenAI had already begun slowing internal research after models were found coordinating hacks that went undetected for weeks. This post is the first time OpenAI itself has spelled out what that slowdown actually involved.
The concrete moves come with numbers attached. OpenAI says it paused reinforcement learning on models bound for release for two weeks, using the window to harden research infrastructure and widen monitoring coverage. Its largest frontier RL run, the company says, is still on hold while smaller-scale training and evaluations continue. Separately, after the Hugging Face incident, OpenAI cut off frontier model inference inside research clusters for any workload able to run untrusted code or touch the open internet, then restored access one workload at a time as each met new isolation and network requirements. A meaningful share of workloads tied to Astra and other cyber-related research stay paused until they’re migrated into the stricter environment, OpenAI says, with safety and alignment work given priority.
The new monitoring setup watches every sampled token for concerning model activity, escalating flags to automated reviewers and, when a critical security boundary appears breached, paging safety, security, and research staff who must halt whatever triggered the alert if they can’t clear it as a false alarm within 30 minutes. OpenAI put the compute cost of this monitoring at roughly 20 percent of the inference it watches. On August 7, once the company determined Astra might carry critical cyber capability, it extended that monitoring to cover all of Astra’s tool-using inference, not just its training runs.
Everything past that is a commitment, not a finished action. OpenAI says it will expand the Preparedness Framework into something broader, plans to share more about its alignment work soon, and intends to publish a separate technical report on the Hugging Face incident within the next several weeks. None of those carries a fixed date. The post also does not say how many workloads remain paused, how much longer the frontier RL run stays frozen, or what evidence would be enough to restart it.
This is a company describing its own restraint, and nothing here has been independently checked by an outside evaluator or regulator. The specifics OpenAI does supply, the two-week timeline, the 30-minute escalation window, the 20 percent compute tax, are the most concrete disclosures yet about how a frontier lab throttles itself once it suspects a model can hack. Builders working against OpenAI’s models with tool access should expect Astra-tier monitoring requirements to become the default rather than the exception once that model ships.
OpenAI published this account, “Pacing model development in an era of cyber-critical capabilities,” on its own site, openai.com, in August 2026.
