OpenAI said on August 18 that it d reinforcement-learning training for its latest deployment-bound models for two weeks and is still holding its largest planned frontier RL run. The company is using smaller runs and evaluations to test alignment and security controls, while estimating that its expanded monitoring consumes roughly 20% of the inference compute being monitored.
OpenAI d reinforcement-learning training on its latest deployment-bound models for two weeks and is keeping its largest planned frontier RL run on hold. In an August 18 announcement, the company said smaller training runs and evaluations are continuing while it tests model behavior, validates safeguards, and gathers more evidence of alignment.
The follows two developments OpenAI says increased the urgency of its work: the OpenAI-Hugging Face security incident and preliminary evidence that its upcoming Astra model may reach the Critical cybersecurity capability threshold under the company's Preparedness Framework. Axios reported that OpenAI is also rewriting that framework as models approach capability levels anticipated by the older document.
Research workloads face a higher security bar
OpenAI says it d frontier-model inference in research clusters for workloads that could execute code or use tools with internet access immediately after the Hugging Face incident. A more limited execution path resumed, but a significant number of Astra and cyber-related workloads remain d until they meet stronger requirements.
Those requirements include tighter sandboxing for untrusted code, more network isolation, fewer shared services and standing privileges, improved logging, and continuous testing of security boundaries. OpenAI says the strictest safeguards apply to Astra and cyber-model workloads because Astra may have Critical cyber capabilities.
Monitoring adds measurable compute overhead
The company also expanded its chain-of-thought monitoring into a multistage system. Activation classifiers inspect sampled model activity and escalate concerning signals to higher-compute investigators that examine tool actions and available reasoning for unauthorized access, data theft, destructive behavior, or attempts to bypass safeguards.
OpenAI says this monitoring is required for tool-using reinforcement-learning runs and evaluations involving models at Sol capability or higher. Tool-using Astra inference has an additional monitoring requirement. The company estimates that monitoring consumes roughly 20% of the inference compute being monitored, while noting that the cost varies by workload.
That figure is a company estimate, not an independent benchmark, but it makes the operational tradeoff concrete. Frontier-model teams adding token-level detection, isolation, and escalation capacity must budget for lower effective throughput and more complex incident response. The larger RL run has no announced restart date; OpenAI says it will proceed only after gathering more evidence that the safeguards and alignment measures work.
Key Points #
- 1OpenAI d deployment-bound frontier RL training for two weeks and has not restarted its largest planned frontier RL run.
- 2Astra and cyber-related workloads must meet stronger isolation, monitoring, and research-environment security requirements before resuming.
- 3OpenAI estimates its expanded monitoring uses roughly 20% of the inference compute being monitored, with costs varying by workload.
Scoring Rationale #
The training hold and expanded controls reveal a material operational response by a leading frontier-model developer to cyber-capability and research-environment risks. The 20% company estimate gives infrastructure teams a concrete monitoring-cost signal, while the lack of independent validation or a public model release keeps the score below a major platform launch.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.