OpenAI has slowed development of its most advanced artificial intelligence models after an unprecedented security incident involving Hugging Face and evidence that an upcoming model could reach what the company considers a critical level of cybersecurity capability.
The AI developer had temporarily slowed the pace of scaling, including a two-week pause in reinforcement learning training on its latest models intended for deployment, while it strengthened monitoring, alignment and security across its research environments.
OpenAI said the decision was driven by two developments: the OpenAI-Hugging Face security incident and preliminary evidence that upcoming model Astra could meet the Critical cybersecurity capability threshold under its Preparedness Framework.
Hugging Face incident raises alarm
The Hugging Face incident occurred during an internal OpenAI cybersecurity evaluation designed to test how effectively advanced models could pursue complex attack paths.
OpenAI said one attack path involved stolen credentials, zero-day vulnerabilities and a remote-code-execution route on Hugging Face servers.
Hugging Face detected and contained the activity, while OpenAI subsequently worked with the company and outside security specialists to investigate the incident.
OpenAI described it as an “unprecedented cyber incident”, saying it demonstrated that advanced AI systems could identify and exploit previously unknown attack paths in real-world systems without access to source code.
Astra adds to cyber concerns
The Hugging Face episode was followed by another warning sign.
OpenAI said on August 7 that internal evaluations of Astra, an upcoming model not involved in the Hugging Face incident, showed substantial improvements in agentic coding and cybersecurity.
Those results meant OpenAI could no longer rule out Astra reaching its Critical cybersecurity capability threshold.
The company has since imposed stricter controls around Astra, including isolated testing environments, restricted network and tool access, stronger model-weight protections, encryption and enhanced monitoring.
Some activities involving Astra have remained paused until those strengthened requirements can be met.
Frontier training deliberately slowed
Following the Hugging Face incident, OpenAI also paused frontier-model inference in research clusters where models could execute code or use tools capable of reaching the internet.
Its monitoring systems now alert safety, security and research teams when potentially concerning behaviour is identified, with activity expected to be paused if the alert cannot be confirmed as a false positive within 30 minutes.
The safeguards apply to reinforcement learning training and evaluations involving tools for models at Sol capability or higher, while all Astra inference involving tools is subject to additional monitoring.
OpenAI estimates that the monitoring itself adds roughly 20% to the inference compute of affected workloads.
The slowdown marks a significant shift in frontier AI development, with OpenAI effectively accepting additional cost and delays to model training as it attempts to keep security protections ahead of rapidly advancing cyber capabilities.