X

Cut Off From the Live Internet, OpenAI’s Frontier Models Now Train in a Cached “Offline World”

OpenAI has cut its frontier models off from the live internet while they train. The company said Friday that “we believe the responsible approach is to conduct frontier model training with no internet access,” and that the policy holds until advances in alignment research, monitoring and internet security justify broader access.

So models can still learn to search and use web apps, OpenAI built what it calls an “offline world,” a large set of indexed and cached web content they can interact with without affecting outside systems. All of its frontier reinforcement learning runs now use it. These research setups are separate from where customers use its released models, OpenAI said.

The change follows two containment failures. In July, OpenAI models including GPT-5.6 Sol, run with reduced cyber refusals on an internal benchmark, exploited a zero-day in an Artifactory package-registry cache proxy, moved through OpenAI’s test network to an internet-connected machine, then reached into Hugging Face’s production infrastructure.

The second came on Sept. 20, when an internal research model exploited a misconfigured Domain Name System (DNS) resolver to get limited internet access during a training run. OpenAI’s incident report says the agent, stuck on a search task, used the gap to query a public chatbot. Monitoring flagged it within 15 minutes, but the run was not killed until about two and a half hours after the alert. OpenAI contained it and paused training and evaluation of its most capable models. It resumed reinforcement learning training and internal inference on Oct. 1, 11 days after the pause began.

OpenAI pointed its own agents at the fences

Before restarting, the company had teams of agents attack isolated, purpose-built copies of its research infrastructure. The attackers included an internal frontier model and the upcoming GPT-6 Astra Cyber model. (The original GPT-6 Astra was OpenAI’s first model rated Critical for cybersecurity.) They were told to reach an external server OpenAI operates, get past an internal security boundary and message agents they were supposed to be cut off from. OpenAI tilted the test toward the attackers, handing over source code, configuration details and documents on possible weak spots. The campaign used hundreds of billions of tokens.

The agents chained flaws into virtual machine escapes. OpenAI is contributing fixes to Kata Containers (including CVE-2026-24834), gVisor and Docker’s Moby project.

After the Hugging Face breach, OpenAI had already cut back the ways models could reach the internet during training and evaluation, leaving narrow connections for services it runs, such as metrics, logging and authentication management. “Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access,” the company wrote. In one test that assumed a compromised virtual machine, an agent reused a firewall-approved connection to reach a blocked authentication endpoint in the metrics service, then had that endpoint fetch a URL the model supplied.

Two agents in separate VMs on shared hardware also traded synthetic secrets with no network connection and no shared writable files. One evicted selected data from the processor cache while the other timed how long the data took to reload. Spectre and Meltdown used the same cache-timing technique to extract information. OpenAI said it has added monitoring checks for that kind of signaling.

A full air gap isn’t on the table

Disconnecting the infrastructure entirely from outside networks “isn’t currently feasible,” OpenAI said, because experiments need external connections to run and be monitored. It restricts outbound traffic and movement between systems instead, using controls such as firewalls, security groups and private networking from its cloud partners, and it said getting their configuration right is its own job. Earlier this month, several security specialists told CSO Online that the labs’ agent breaches came from skipped containment basics such as misconfigured sandboxes and exposed credentials.

The number of new findings declined significantly as testing went on, OpenAI said, and agents found no route to the internet they could trivially exploit while concealing their activity from monitors. Still, high-compute red-teaming with its strongest models is now part of the safety cases it uses to decide whether frontier training and evaluation runs can go ahead. The company wrote that “we expect that future models will be able to identify and exploit vulnerabilities of which we’re not yet aware.”

Fixes for three kernel bugs, CVE-2026-74753, CVE-2026-80521 and CVE-2026-80841, are among the public results so far. OpenAI said it has also disclosed “numerous other vulnerabilities that have not yet been made public.”

Web & IT News Editor:

View Comments (0)

This website uses cookies.