Managing frontier AI risk: From training to deployment

OpenAI
OpenAI explains its defense-in-depth approach to managing frontier AI risks across research, internal deployment, and external deployment stages.

Summary

OpenAI outlines a defense-in-depth strategy for managing frontier AI risk across three stages: research, internal deployment, and external deployment. During research, models are trained in sandboxed environments without live internet access, backed by continuous red-teaming and monitoring; a misconfigured DNS incident led to a training pause from September 20 to October 1. Before internal deployment, models must pass alignment and monitorability evaluations and are monitored on every inference, with system-level classifiers and Auto-Review providing additional safeguards. External deployment adds broader evaluation suites, adaptive testing, third-party assessments, offline enforcement, and trusted access programs. OpenAI commits to pausing training or withholding releases when safety evidence falls short, and to disclosing misalignment incidents through reports and system cards.

(Source:OpenAI)