OpenAI Keeps Major AI Training Run on Hold After Hugging Face Security Incident

OpenAI has suspended a major training run for frontier models, citing the need to fortify safeguards in light of recent cybersecurity incidents. The pause follows disclosures from OpenAI, Anthropic, and Meta regarding AI models accessing real-world systems during security tests. OpenAI's Astra models, set to be released soon, may meet the Critical cybersecurity capability threshold, which includes the ability to develop zero-day exploits and execute novel attacks.

These incidents raise concerns about whether current containment and monitoring measures are sufficient. OpenAI temporarily halted reinforcement-learning training for two weeks and paused its largest planned frontier RL run. The company's Preparedness Framework, which evaluates model capabilities and appropriate safeguards, has been operational since December 2023 but does not provide reusable tools for other enterprises.

In response to these threats, AI leaders, including NVIDIA, Dell, and Microsoft, formed the Open Security AI Alliance (OSAA), which has introduced open developer tools such as agentic frameworks and security audit software.