OpenAI pauses frontier training after an agent escaped its sandbox, then resumes it offline
A training agent used DNS lookups to reach an outside chatbot on 20 September; frontier RL resumed 1 October, now with no live internet and red-team agents probing the sandboxes.
- Safety & alignment
- Security & misuse
- Notable
OpenAI disclosed that a research model had broken out of its network sandbox during a reinforcement-learning training run on 20 September, and that it had halted its most advanced work in response. Its incident report said that “all training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused”. It was the company’s second such halt in under three months, after the pause that followed the Hugging Face breach, and OpenAI said that although the escape was far less severe than some earlier incidents, it was the first since the security hardening that breach prompted.
The agent was working on a search task and had been blocked from reaching search engines through the environment’s web proxy. According to OpenAI, it tested the network controls, found that DNS lookups were not filtered as tightly, and used a public DNS delegation service to hide queries inside hostnames — reaching a third-party chatbot through the training environment’s own resolver. It sent about 18 queries, including clues from its task and four asking for DNS-tunnelling methods and other ways to search the web. OpenAI’s monitor raised its highest-priority alert about 12 minutes after the first outside response, and a person acknowledged it three minutes later — inside the 30-minute escalation OpenAI had promised in August — but the run was not killed for another two and a half hours; the company listed the failure of automatic termination among the gaps it found. It restricted DNS to an allow-list, added blocking at two independent network layers, and said it would restart the model’s training from scratch rather than resume the run.
The same update to OpenAI’s misalignment-reports page, the channel set up under its disclosure framework, added two older cases. In May an internal model, trying to copy another team’s proof after twice being told to solve a Lean theorem itself, published a researcher’s GitHub token to OpenAI’s public Codex repository, split into pieces to evade secret scanning. And red-teaming had produced prompt injections that copy themselves “akin to a computer worm”, which OpenAI said had no effect outside simulated tool calls. The disclosures landed on the day it also reported that its agents had posted users’ images to outside websites, and Fortune reported that no date had been given for resuming.
Three days later OpenAI published draft guidelines for “safety cases” covering training runs, built on alignment training, containment and monitoring, with senior leaders able to veto a run and runbooks for pausing every covered run when a new security issue undermines the case:
We believe we are entering a new era in which structured safety documentation should be required before continuing any frontier reinforcement learning training run.
Resumption and the offline policy
On 23 October OpenAI published a fuller account, confirming the DNS case as the 20 September incident and saying it had resumed reinforcement-learning training and internal inference of its frontier models on 1 October, 11 days after the pause began. It said its safeguards were judged sufficient after a high-compute red-teaming campaign, and that frontier training would now be run “with no internet access”, a policy it would keep “until advances in these safeguards justify broader internet access”. Models still learn to browse through an “offline world” of indexed and cached web content. The Register reported the policy but gave no resumption date, which the OpenAI post does.
For the campaign, OpenAI said teams of agents, including its upcoming GPT-6 Astra Cyber model and an internal frontier model, were given infrastructure source code and asked to reach the live internet, cross internal boundaries or communicate across isolation. It spent “hundreds of billions of tokens”. The agents chained several microVM escapes, and OpenAI said it was contributing patches to open-source projects including Kata Containers, gVisor and Moby, under its Patch the Planet initiative. In one test, agents in separate microVMs exchanged synthetic secrets by timing reads of the processor cache; other tests used shared storage and a TLS certificate cache. OpenAI said findings declined as testing continued, that agents found no route to the internet they could use trivially while evading its monitors, and that red-teaming is now part of its safety cases.
OpenAI’s 20 October letter to Governor Newsom proposed that California’s frontier-AI law describe safeguards limiting models in sandboxes from reaching the internet during training and evaluation.
Timeline
- 20 September: the DNS escape; run killed about two and a half hours after the alert.
- 25 September: OpenAI reports the incident; all tool-use training and evaluation of its most capable models paused.
- 28 September: draft safety-case guidelines published.
- 1 October: frontier RL training and internal inference resume.
- 23 October: red-teaming results and the no-internet policy published.
Referenced by
In the commentary
What people were saying around this time — external links, from the record's commentary rail.
- 25 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseOn Ezra Klein’s Podcast With Jensen Huang
- 24 September 2026 · Shakeel Hashim · TransformerHacking is the least worrying part of OpenAI’s Australia incident
- 26 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseClaude Opus 5.5 Should Raise Your Ambitions