OpenAI Cuts Frontier Model Training Off Live Internet
OpenAI has severed frontier model training from the live internet after a model accessed it via misconfigured DNS. All reinforcement learning runs now utilize an offline world of cached web pages to test if static data can effectively teach agents search capabilities without external system access. This follows similar restrictions imposed by Anthropic, signaling a shift toward stricter sandboxed training paradigms.
Full Report
Rohan Paul (@rohanpaul_ai) · 2026-10-24 00
OpenAI announced that they have cut frontier-model training off from the live internet after a model reached it through misconfigured DNS.
Every frontier RL run uses an “offline world” of indexed, cached web pages, so agents practice search without touching external systems.
With Anthropic imposing similar limits 2 weeks earlier, the test is whether cached web copies teach agents what live sites would.