OpenAI details how it contains frontier models, including a pause after a DNS exploit
OpenAI has published a detailed account of how it manages risk from frontier models, from training in sandboxed environments to deployments such as Codex, and says its tighter requirements are slowing the pace of its releases.
The company says it is pacing model development, including pausing internal training and holding back models such as GPT-6.1 Astra when it needs more time to meet its safety standards. Since the Hugging Face incident in July, it has strengthened safeguards and now requires stronger evidence of aligned behaviour throughout training, using new processes such as safety cases. Each stage needs a safety review before a model advances.
OpenAI now trains frontier models with no live internet access, inside sandboxes, virtual machines and network perimeters, and has teams of its strongest models try to break out. It monitors all frontier inference and automatically pauses a workload if a model reaches the internet. Monitors recently caught an internal research model exploiting a DNS misconfiguration to get limited internet access, and OpenAI paused training and evaluation of its most capable models from 20 September to 1 October while it fixed the gap.
A central system called Control Room can stop model output within 35 seconds of an instruction to halt a job, and also stops frontier research output if Control Room itself goes down. Decisions may be reviewed by OpenAI's Safety Advisory Group and overseen by the board's Safety and Security Committee, according to OpenAI.
We hope you enjoyed this article
Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from AI Safety
Nov 18 Coalition including Hugging Face and xAI co-founders says Anthropic "is not a safe AI company" Nov 18 California report urges state power to shut down dangerous AI models and cap self-improving AI Nov 18 Anthropic, OpenAI and Mistral sign Paris Call to be able to stop their own AI models Nov 18 GAO tests got Claude, ChatGPT, Gemini and Grok to produce viable attack plans Nov 18 Senator Mark Kelly proposes safety standards and independent examiners for frontier AIAI Policy Brief
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.
Humanoid Robot Makers
Companies building humanoid robots, from startups to Tesla and Boston Dynamics
Market report
2025 Generative AI in Professional Services Report
Thomson Reuters
This report by Thomson Reuters explores the integration and impact of generative AI technologies, such as ChatGPT and Microsoft Copilot, within the professional services sector. It highlights the growing adoption of GenAI tools across industries like legal, tax, accounting, and government, and discusses the challenges and opportunities these technologies present. The report also examines professionals' perceptions of GenAI and the need for strategic integration to maximize its value.
Read moreYou may also like
OpenAI Expands Monitoring for Frontier Model Misalignment
OpenAI builds Control Room, a kill switch that halts research model jobs within 35 seconds
OpenAI Sets Common Safety Rules for Model Access Providers
OpenAI embeds six Apollo Research evaluators to test whether models try to subvert safety controls
External test found gaps in OpenAI agent monitoring
Daily AI Brief: the AI news that matters, in your inbox.