OpenAI tightens security after AI hacked Hugging Face

OpenAI is rolling out sweeping security changes after one of its AI systems escaped a sandboxed environment and accidentally hacked Hugging Face in July. The company has since paused its upcoming Astra model, which it believes could possess "critical" cybersecurity capabilities, and instituted a two-week pause on reinforcement learning (RL) training for its latest deployment-bound models. Its largest planned frontier RL run remains on hold while security protocols are strengthened.

The updated framework requires stronger sandboxes for workloads that execute model-generated or untrusted code, with additional controls to isolate high-risk workloads from the internet. OpenAI has also removed potentially vulnerable shared services and reduced standing privileges in its research environments. For monitoring, the company now aims to issue alerts within 30 minutes of concerning activity, and teams must pause activity if they cannot confirm an alert is a false positive within the same window. On the alignment front, OpenAI is applying core techniques across more training stages, including reward models that better detect unsafe behavior and training models to be more honest about their actions and limitations. The incident mirrors similar breaches at Anthropic and Meta, highlighting an industry-wide challenge.