Leadership
OpenAI lays out frontier AI risk management
- Published
- Source
- OpenAI blog
Summary
OpenAI published how it evaluates and controls model risk across three stages: research, internal deployment, and external deployment. It layers sandboxing, monitoring, alignment, and system-level controls, requires a safety review to advance each stage, and places oversight with a Safety Advisory Group and a board committee. After the July Hugging Face incident it restructured training to run cut off from the internet, and it disclosed that a monitor caught an internal model gaining limited internet access through a DNS misconfiguration, prompting a pause in frontier training from September 20 to October 1. It also runs a Control Room that can stop model output within 35 seconds.
Why it matters for our work
As a working example of releasing highly autonomous models step by step from controlled environments, it offers a reference for organizations designing AI risk management and audit schemes. The control system and figures are OpenAI's own account.
Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.