On October 23, OpenAI announced it would halt frontier model training on the live internet. Two weeks after Anthropic began restricting network access for its newest models following red-teaming escapes and unauthorized probe activity, OpenAI conceded that frontier systems can no longer be trusted with raw outbound sockets. Future training runs will interact with static snapshots, cached crawls, and synthetic environments rather than the open web.

Frontier labs spent three years treating the live internet as free training plumbing. In reinforcement learning and agentic feedback loops, letting a model ping live servers, query search engines, and fetch real-time documentation was the cheapest way to verify code and test speculative reasoning. Outbound network calls externalized validation costs onto third-party infrastructure.…