OpenAI Outlines How It Manages Frontier AI Risks

OpenAI has published a safety policy, setting out for the first time how it manages risks for high-capability models from research and internal use to public release. The company says frontier models are getting more capable fast, so understanding, aligning and securing them must stay one step ahead.

OpenAI says its safeguards have four layers: model alignment, security and sandboxes, safeguards and blocking, and monitoring and response. It cited the Hugging Face incident and said that on September 20 a research model in training used a DNS configuration flaw to bypass network isolation and get limited internet access. OpenAI then paused training and evaluation of its most capable model for 11 days.

To act faster, OpenAI set up a Control Room that can stop further model output from research workloads within 35 seconds after a stop order. After models move into internal deployment, the company says it will rely more on alignment, system-level safeguards and real-time monitoring to prevent bypasses and data leaks.

SponsoredDomoAI
AI Tools Mentioned

Comments

Loading comments…