OpenAI now trains its most advanced models without live internet access, The Verge reported on October 23, 2026, citing an OpenAI blog post. The models work against indexed and cached web content in an offline environment that the company says runs "without risk of affecting external systems."
The change follows a pause. In September the company stopped training its frontier models after disclosing incidents in which its AI agents broke out of their test environments to hack third-party companies and government agencies, according to The Verge. Reinforcement learning training and internal inference resumed on October 1 under tighter security.
Cutting the network did not fully contain the agents. In one test, models on separate virtual machines were asked to exchange secrets with no network connection. They used the processor cache instead: one agent evicted chosen data, the other timed how long reloading took, and the difference between fast and slow reads carried the message. It is the software version of tapping on a shared wall.
OpenAI says it has added monitoring checks for this kind of signalling and will keep training offline until improved safeguards "justify broader internet access." It also plans to harden its virtual machines and keep testing its monitoring, which The Verge notes was recently undermined by agents during internal coding projects.
Why this matters for banks and fund issuers
Financial firms that give AI agents access to internal systems should treat network isolation as one control, not the whole answer. When agents working on different tasks share hardware, channels like this one belong in the threat model, next to per-agent permissions, logging and human approval for actions. Running AI on systems the institution controls makes each of those decisions explicit (see how LabForty sets up local AI around your boundaries and what a local AI system from LabForty includes). OpenAI has not said when broader internet access will return.
