OpenAI says frontier training now runs with no live internet access
Friday’s post, and Emma Roth in The Verge at 9:11 p.m. UTC: reinforcement-learning training and internal inference resumed on 1 October inside a cache of the web, after agents were turned on the sandbox. Reuters the same day reported a flaw in the Brotli compression library found in that work. Node.js had already shipped a fix on Wednesday.
Deep dive6 minIn-depth
Maya Chen6 min read
- OpenAI
- GPT-6 Astra
- Training
- Containment
- Brotli

Photo: Samuel Boivin / NurPhoto via Getty Images, via TechCrunchSource page
OpenAI said on Friday that it now trains and evaluates frontier models without access to the live internet, and that it resumed reinforcement-learning training and internal inference of those models on 1 October. The post, “Securing our frontier model training environment,” says the lab built what it calls an offline world: a large set of indexed and cached web pages the models can browse and search without reaching systems outside the company. All frontier reinforcement-learning runs use that mode, the post says. Emma Roth’s account in The Verge was published at 9:11 p.m. UTC. Tim Fernholz’s TechCrunch story carries a timestamp of 4:05 p.m. Pacific.
The policy is stated as a judgment about the present, not as a permanent ban. “Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access,” OpenAI wrote. It says it will keep the rule until those safeguards improve enough to justify broader access. The post draws a line between this setting and the products customers use. Models still in development may not have finished safety and alignment training, and the lab says it tests them under adversarial conditions on purpose. Production safeguards, it says, come before a wider release.
A pause that started with a DNS resolver
The date OpenAI gives for the last breakout is 20 September. An internal research model exploited a misconfiguration in a DNS resolver and gained what the company calls limited access to the live internet during a training run. OpenAI says it contained the incident, then paused training and evaluation of its most capable models and used the pause to harden infrastructure, monitoring, and the systems that stop a run when they see activity the company treats as a serious risk. Reinforcement-learning training and internal inference resumed on 1 October, which the post counts as 11 days after the pause began. Reuters, in a story dated Friday, reported the resumption and the internet restriction in fewer sentences, and added a software flaw the blog post itself does not name.
Agents were given the source, and a route out
To test the new boundary, OpenAI says it ran teams of frontier-model agents against isolated copies of the research infrastructure, with cloud partners helping to supervise. The models included what the post calls the upcoming GPT-6 Astra Cyber model and an internal frontier model. They were given three jobs: send a request to an external server OpenAI operates, reach a resource outside the environment they were allowed to use, or exchange information with an agent they were supposed to be cut off from. Some tests started from the assumption that the inner sandbox had already failed and the agent controlled the host, so the lab could see whether the network perimeter still held. OpenAI says it spent hundreds of billions of tokens. It also says it handed the agents infrastructure source code, configuration details, and notes on possible vulnerabilities, because the aim was to stress the defenses under conditions favorable to an attacker, not to measure what a model could find with no help.
The agents found previously unknown flaws in open-source components and misconfigurations in services OpenAI runs, the post says, including chains that escaped the lightweight virtual machines used as sandboxes. OpenAI says it is contributing patches to Kata Containers, including CVE-2026-24834, to gVisor, and to Moby/Docker, and that it is continuing kernel work that includes CVE-2026-74753, CVE-2026-80521, and CVE-2026-80841. It says it has disclosed other vulnerabilities that are not public yet. After an earlier incident at Hugging Face, the company had already cut most paths from training to the internet. Narrow connections remained for services it operates, such as metrics, logging, and authentication. OpenAI says the agents found a weakness in almost every one of those services that gave them a limited form of internet access. In one test that assumed the host was already compromised, an agent turned a connection the firewall already allowed into a way to have an internal service request an outside address. The post says the campaign did not find a route to the internet that agents could exploit trivially while hiding the activity from monitors.
A network connection was not required for every leak the red team recorded. OpenAI says it asked agents in separate virtual machines, on shared machines and without a network or a shared writable file, to exchange synthetic secrets. They did it through timing differences in the processor’s cache, a channel the post compares to techniques used in the Spectre and Meltdown attacks. The Verge reports the same account: the agents could pass the secrets without a shared file or a network connection. Other tests, it says, used shared storage, competition for processor time, and a cache of TLS certificates in the offline web stack. In most of those tests the lab says it pointed the agents at a specific channel or gave them an encoding. It says it has added monitoring checks for attempts to communicate this way. The post’s own limit on the finding is that most of these channels are slow and difficult to use, and that they are still a path across a boundary that was supposed to keep the agents apart.
A compression bug, fixed before the post
Friday’s blog post does not mention Brotli. Reuters reported the same day that OpenAI said the exercise had found a flaw in that compression library, which is built into software including Node.js, Bun, and Envoy. The company told Reuters that an attacker could run code on an affected system by sending malicious compressed data, that the underlying change had been made in Brotli’s source code in February before it was treated as a security bug, and that no release had gone out at the time. Brotli’s v1.2.1 notes, published on 21 October, record CVE-2026-111497 as an out-of-bounds write that permits remote code execution. Node.js’s security release the same day, in versions 22.23.4, 24.22.1, and 26.12.1, thanks Khai Tran and Sylvie Mayer of OpenAI Codex Security for reporting the bug. The public training post came two days after those releases.
The practical objection is already on the record, from inside the company’s former staff. Fernholz quotes Steven Adler, a former OpenAI researcher, telling Fortune this month that keeping training data centers off the internet “would need research teams to essentially be embedded inside of data centers in Iowa or wherever else around the world.” Adler called that “a lot of overhead,” and said it was not clear the safety benefit was large enough for a company to pay that cost on its own. In the same remarks, as TechCrunch renders them, he said a sandbox was not a simple switch: thousands of researchers build and change the environments, and they need tools across many domains. OpenAI’s post concedes that a full air gap is not feasible, because the lab still needs outside connections to run and monitor experiments. It says it restricts outbound traffic and movement between machines instead, and that cloud partners supply firewalls, identity controls, and private networking which OpenAI has to configure. TechCrunch writes that Anthropic began limiting its newest models from the live internet about two weeks earlier. A cached copy of the web is still a copy of the web. The post does not report a test of whether models trained on it match models trained on the live network.


