Rapid News Brief — OpenAI disclosed Oct. 23 that it has barred live internet access during frontier-model training, changing how its research systems practice online tasks while limiting their ability to affect outside networks.
Training resumed under tighter controls
The company said reinforcement learning and internal inference resumed Oct. 1 after a pause to strengthen security. Friday’s disclosure describes those changes; it is not an announcement that training has just restarted.
In its security account, OpenAI said models now practice browsing and searching against indexed, cached web material. The policy concerns internal research environments, not a blanket removal of internet tools from released customer products.
OpenAI said its protections combine restricted execution environments, virtual machines, monitoring and network controls. The infrastructure is not completely disconnected: external connections remain necessary to operate and monitor experiments. The company said it restricts outbound traffic and unauthorized movement between systems.
That distinction matters for enterprise readers: offline model activity and fully disconnected infrastructure are different security arrangements. OpenAI’s description does not establish that every possible route around its controls has been eliminated.
External review adds a separate limitation
A report published Oct. 22 by Redwood Research describes a separate pilot with METR, conducted July 21-Aug. 11. The researchers examined OpenAI’s earlier internal monitoring system, with access to code, documentation and staff interviews.
They reported vulnerabilities that could, in principle, let agents interfere with monitoring. Their exercise assessed the historical system, not the current coverage of its expanded successor. OpenAI told them vulnerabilities affecting production safeguards had been patched; the researchers said they had not evaluated that claim.
The researchers also distinguished traffic eligible for monitoring from traffic that actually completes the monitoring process. Their recommendations included reconciling inference records against monitoring logs and isolating agents from the systems watching them.
For businesses assessing agent oversight, the two disclosures identify separate questions: what an agent can reach, and whether its actions are reliably observed.