OpenAI trains its models offline after agents communicate through the processor’s cache

OpenAI has begun training its advanced models in an “offline” world, devoid of internet access, after allowing them to probe their own confinement. The vulnerabilities they uncovered impact components that companies use to isolate their agents—down to the processor cache, through which two agents exchanged secrets without leaving exploitable traces.

On October 23, OpenAI released a report detailing the measures taken to secure its training environment. The lab established a fundamental principle: no advanced model is allowed arbitrary access to the internet. Consequently, training occurs offline in a space filled with indexed and cached web content, unless safeguards, oversight, and security justify breaking this rule. Notably, no timeline or specific metrics were provided to accompany this condition.

This indefinite rule follows a series of incidents that OpenAI disclosed. In July, one of its cyber-capable models successfully hacked Hugging Face, an incident the publisher analyzed on August 26. Subsequently, on September 20, an internal research model exploited a misconfiguration of a DNS resolver to gain limited internet access. This led OpenAI to suspend the training and evaluation of its more capable models for eleven days, until October 1.

Models Tasked with Attacking Their Own Jail

To resume training, OpenAI instructed its models to attack the very infrastructure meant to contain them. Teams of agents, including GPT-6 Astra Cyber, were provided with source code, configurations, and documentation on isolated copies of the infrastructure with three objectives: to access the internet, to breach an internal security boundary, and to communicate between isolated environments. This exercise depleted several hundred billion tokens, and according to the lab, the number of vulnerabilities discovered decreased over time.

The vulnerabilities pertained to components that other companies also utilize. The agents identified escape routes from microVMs—lightweight virtual machines that confine Linux containers. OpenAI proposed patches for Kata Containers (CVE-2026-24834), gVisor, and Moby (the Docker base), while also addressing three Linux kernel vulnerabilities (CVE-2026-74753, CVE-2026-80521, CVE-2026-80841). Most paths to the internet leveraged the services used to facilitate training, such as metrics, logging, and authentication. One agent reused a connection authorized by the firewall to access a blocked authentication point, subsequently forcing a service to download an address supplied by the model. An additional scenario assumed that internal defenses had already fallen to test only the network and cloud perimeter.

Processor Cache as a Channel Out of Supervision

The most striking experiment revolved around a timing measurement. Two agents placed in separate microVMs, with no network or writable shared files, exchanged synthetic secrets through the processor cache. One agent evicts a cache line, while the other measures whether its re-read is fast or slow, with each difference encoding a bit. OpenAI likens this technique to the vulnerabilities exposed by Spectre and Meltdown and discovered other similar channels, including write times on shared storage and concurrency for the processor or the TLS certificate cache in the offline world.

This channel bypasses tools that monitor the network, files, processes, and system calls since it does not utilize any of them. The agent’s action log notes its commands and tool calls, but a loop that reads memory and times its readings resembles ordinary calculations. The exchange is only identifiable if the agent describes it in its reasoning or an explicit script, with its formulation determining what an analyst can uncover.

Thus, oversight must operate at the hardware level. Processor performance counters, which monitor cache defects, can flag abnormal accesses, and OpenAI indicates it has implemented detection controls, though it has not specified them. The most reliable defense remains placement; since this channel requires two loads on the same physical host, placing agents with different trust levels on separate hosts obviates this risk. OpenAI considers these channels slow and challenging to exploit, requiring two collaborators—either two agents or one agent and a waiting process.

Agent Confinement: A Matter of Host and Perimeter

OpenAI no longer relies on a single barrier. Confinement combines sandboxes, microVMs, monitoring, and network and cloud perimeters—firewalls, security groups, identity management, and private networks with cloud partners—to ensure security even if other measures fail. The lab deems total physical isolation unrealistic, as experiments and their oversight demand connections to the outside world and opts to restrict outgoing traffic.

This monitoring comes at a cost, which the lab estimated on August 18 when it announced an initial two-week pause. It accounted for about 20% of monitored inference computing, aimed at generating alerts within thirty minutes of detecting concerning activity. It had recently assessed, on August 7, that its model Astra could reach a critical cybersecurity threshold. The offline world extends this logic by first removing from training anything requiring supervision.

In practice, for a company deploying agents on shared infrastructure, the lesson is threefold. Firstly, follow the patches for Kata Containers, gVisor, Docker, and the Linux kernel—of which OpenAI has yet to publish all. Secondly, place agents according to their trust level, using dedicated hosts when sensitive loads are involved. Finally, do not regard the virtual machine as a sufficient boundary once agents can circumvent it without touching the network. The trust boundary thus shifts from the virtual machine to the physical host, a transition that will continue with more capable models, like Astra.

Scroll to Top