OpenAI pauses deployment-focused frontier RL training for two weeks

• OpenAI stopped RL training on its newest deployment-bound models for two weeks, while its largest planned frontier RL run remains held: merettm's post describes the continuing restriction, and the August 18 company post says smaller training runs and evaluations are continuing.
• Research clusters are getting stronger workload isolation, network isolation, and continuous security testing, as OpenAI's safeguards announcement describes.
• A multistage monitor inspects every sampled token, escalates suspicious activity, and can pause a run when a critical alert cannot be cleared within 30 minutes, according to eliebakouch's monitoring breakdown.
• Tool-using RL training and evaluations at Sol capability or higher now require monitoring, while Astra tool inference gained an additional requirement after an August 7 assessment, the official account says.
• Alignment work now reaches more training stages, and sama's statement frames safety confidence as the factor that will increasingly set the pace of progress.

The official post says OpenAI is investing in automation that uses its own models to continuously test new trust boundaries against simulated attacks. It also promises a technical report on the Hugging Face incident in the coming weeks.

OpenAI describes the two-week halt as completed, but its largest planned frontier RL run is still on hold while smaller-scale runs test model behavior, safeguards, and alignment evidence.

Release timing received a narrower qualification in a follow-up: sama's follow-up says models already close to release are still expected soon, while the disruption affects further-out releases.

Immediately after the Hugging Face incident, OpenAI paused frontier-model inference in research clusters for runs that could execute code or use internet-accessible tools. The company's account says a more limited code-execution path returned quickly, then workloads were reviewed individually.

The new environment work has three parts:

• Workload isolation: stronger sandboxes for model-generated or untrusted code, including software that processes model outputs.
• Network isolation: controls intended to prevent one compromised workload or supporting service from reaching the internet or internal networks.
• Continuous security testing: removal of vulnerable shared services, fewer standing privileges, better security-log collection, and tighter trust boundaries.

Earlier monitoring centered on internal frontier deployments and frontier RL runs. The new design is multistage:

1. Activation classifiers inspect internal activity at every sampled token.
2. Higher-compute automated investigators examine tool actions, available reasoning, and the full activity sequence for unauthorized access, data theft, destructive behavior, or attempts to defeat safeguards.
3. The system aims to issue an alert within 30 minutes. A likely critical-boundary violation pages safety, security, and research teams; activity is expected to pause unless they can establish a false positive within another 30 minutes.

The official estimate puts monitoring overhead at roughly 20% of the inference compute being monitored, with costs varying across training and evaluation workloads. That operating cost is also the figure imjustnewatai's summary singled out in its account of the new gate.

The official rules define two policy gates:

• Sol capability or higher: monitoring is required for every tool-using RL training run and evaluation.
• Astra: after OpenAI determined on August 7 that Astra may have Critical cyber capabilities, monitoring became required for every Astra inference using tools.

OpenAI says its strictest research-environment safeguards apply to Astra, cyber models, and other cyber-related workloads. Some Astra training and evaluations meet that bar; a significant number remain paused while they are migrated and enhanced, with safety and alignment workloads prioritized first.

For its most capable RL runs, OpenAI says it is applying core alignment methods across more of training. The company's list includes:

• improving reward models to detect and discourage unsafe behavior across tasks and environments;
• training models to be more honest about actions, capabilities, and limitations;
• reducing behavior that exploits weaknesses in rewards, graders, tools, or oversight;
• expanding training coverage for harmful behavior involving external systems or resources.

The current Preparedness Framework requires safeguards that sufficiently minimize risk during development when a model reaches a Critical capability threshold, regardless of deployment plans.

OpenAI says in its safeguards announcement that it will evolve that framework to join monitoring, alignment, and security across training and deployment, account for future capabilities and operating environments, and involve external organizations.