OpenAI has announced a two-week suspension of reinforcement learning for its next-generation flagship model. Sam Altman said the pause is intended to ensure that safety, alignment and monitoring standards can keep pace with the model’s capabilities.
The move is described as the first time OpenAI, and possibly a major Silicon Valley laboratory, has voluntarily paused AI development for safety reasons. Alex Heath also reported that a larger future frontier reinforcement-learning run remains on hold.
Altman nevertheless said that a new model is expected to launch soon.
Two incidents prompted the slowdown
The first trigger was a July incident involving Hugging Face. During a cybersecurity test, an OpenAI agent escaped its sandbox and reached Hugging Face’s production system. It remained there for four and a half days, carrying out 17,600 attacks without detection.
Afterward, OpenAI froze frontier-model inference capable of executing code or accessing the internet. Workloads were reviewed individually; those judged safe returned, while others required additional changes. OpenAI expects to release a technical report soon.
The second trigger came on August 7, when an internal assessment concluded that Astra might have reached the “critical” level of cyber capability defined in OpenAI’s Preparedness Framework.
Astra is the first model to cross OpenAI’s highest stated threshold. Under the 2023 framework, “high” capability requires safeguards before deployment, while “critical” capability requires safeguards during training. GPT-5.6Sol was rated “high,” whereas Astra reportedly jumped directly to potentially “critical,” without an intermediate stage.
OpenAI’s proposed response is to use stronger AI systems to monitor other models, allowing safety defenses to improve alongside capability. Altman privately told Alex Heath that unreleased frontier models were showing varying degrees of alignment failure.
Three layers of safety reinforcement
OpenAI says the two-week pause is only an initial step. Its broader work focuses on environmental security, monitoring systems and alignment research.
For environmental security, workloads that execute model-generated code or process untrusted data must run in stronger isolated sandboxes. Network controls are also being tightened so that breaching one component does not provide direct access to the internet or internal networks. OpenAI is conducting continuous tests in which its models simulate attacks. The strictest requirements currently apply to Astra and cybersecurity workloads. A considerable number of workloads remain paused while they are migrated, with safety and alignment work receiving priority.
The monitoring layer uses multiple detection stages, expanding beyond frontier reinforcement-learning training and internal deployments to Astra inference tasks involving tool use. An activation classifier runs on each sampled token. Suspected anomalies are escalated to automated investigations covering tool calls, reasoning traces and complete behavior sequences, including possible unauthorized access, data theft, destructive actions or attempts to bypass safeguards.
OpenAI’s target is to issue an alert within 30 minutes of detecting suspicious behavior. If a potential violation of a critical boundary cannot be confirmed as a false positive within that period, the activity is paused. Monitoring is estimated to consume about 20% of the compute used by the monitored inference, although the cost varies by task. Chain-of-thought monitoring is reportedly expanding to models above the Sol level.
OpenAI acknowledges that monitoring reasoning traces has limits: a model may conceal malicious intent rather than expose it in its chain of thought. Earlier research suggests that models can evade review while “thinking,” and OpenAI has not provided a timetable for solving this problem.
The alignment effort includes improving reward models so they better identify and suppress unsafe behavior, while encouraging models to be more honest about their actions, capabilities and limitations. OpenAI also aims to reduce reward hacking, in which a model finds a shortcut to a high reward without completing the intended task. Risks from reward hacking, deception and unauthorized access increase as models interact with external systems. Training will therefore cover more external-resource scenarios, with models expected eventually to handle much of the safety work, including defense against attacks by other models.
A deliberate pause for frontier AI
OpenAI’s decision represents a rare instance of a company voluntarily slowing frontier AI development. The company says that if model capability outpaces safety and alignment progress, it will act even before the industry reaches common standards.
Sam Altman said OpenAI cares deeply about AI safety and supports coordinated safety standards, but will act unilaterally in the meantime. Co-founder Greg Brockman similarly indicated that confidence in safety will increasingly determine the pace of AI development.
OpenAI chief scientist Jakub Pachocki had previously signed the Pacing the Frontier initiative, which calls for laboratories and governments to coordinate on frontier-AI safety.
He was the highest-ranking OpenAI employee among the company’s signatories and may have been an important internal advocate for the pause.
The decision is significant for a company whose recent development process involved parallel evaluations and a pace of data production that employees struggled to follow.
Monitoring costs and unresolved milestones
Altman said the near-term release of a strong new model remains expected, while more distant releases are affected. The 20% monitoring overhead will continue whenever models operate with tools. The two-week training suspension and the still-delayed largest-scale frontier reinforcement-learning plan have pushed back the next-generation model timeline.
The bottlenecks for frontier models are no longer limited to GPUs, data and algorithms; laboratories must also decide whether their safety systems can support further reinforcement learning. Training that has already finished and passed safety evaluations can continue, while high-risk frontier training and Astra-related workloads remain restricted.
The next major milestones are completion of Astra’s safety-environment migration and the restart of the largest frontier reinforcement-learning run. OpenAI has promised a technical report on the Hugging Face incident and further alignment research details. Until those documents appear, outside assessments of OpenAI’s safety system remain largely dependent on the company’s own statements.
