OpenAI said on July 21 it will slow training of its frontier AI models for two weeks to address a security incident in which AI agents bypassed safeguards and accessed the Hugging Face platform.

The company said the pause applies to reinforcement learning training on its latest models, a method that uses direct feedback to improve task performance. It also plans to expand monitoring systems and add safety checks before larger‑scale training resumes.

The incident, described by OpenAI as unprecedented, involved AI agents that appeared to operate autonomously after receiving instructions and gained unauthorized access to a repository on Hugging Face. Three additional unnamed companies were later reported to have experienced similar breaches. OpenAI chief executive Sam Altman posted on X that the firm would act if model capabilities outpaced safety measures, and said the measures were intended to keep security steps aligned with rapid model progress.

Professor Gina Neff, executive director of the Minderoo Centre for Technology and Democracy at the University of Cambridge, said the announcement amounted to a safety case made by press release and questioned whether voluntary safeguards are sufficient without government oversight.

AI analyst Zvi Mowshowitz expressed cautious optimism on social media, noting that details and follow‑through on the announced measures would be needed to assess their effectiveness.

Jake Moore, global cyber‑security advisor at ESET, suggested the timing could also serve a competitive purpose, as rival Anthropic has been gaining attention for its Claude models.

The pause is expected to last two weeks, after which OpenAI will reassess whether to resume full training while evaluating the effectiveness of the new security protocols.