OpenAI says it will no longer train its AI models on the live internet, citing “significant risks”

The company said future training runs will be conducted in an environment that has no direct access to public web resources. The Verge’s piece titled “OpenAI now trains its frontier AI models with no internet access” foregrounds the technical change. Reuters’ article emphasizes that the training restart follows a security overhaul and discloses a flaw in widely used software uncovered during the process. Bloomberg.com connects the decision to recent cyber incidents that affected OpenAI’s infrastructure.

TechCrunch’s story repeats the company’s wording about “significant risks” and frames the move as a response to those threats. The decision represents a marked departure from OpenAI’s earlier approach of continuously pulling data from the open internet to keep its models current. Industry practice has generally favored live‑web ingestion for breadth, freshness and diversity of training material. Bloomberg’s reference to “cyber incidents” and Reuters’ identification of a software flaw suggest that vulnerabilities in the internet‑connected pipeline prompted a reassessment of risk.

By isolating the training environment, OpenAI aims to mitigate exposure to external attacks while still advancing its frontier model capabilities. Going forward, analysts will monitor OpenAI for additional disclosures about the specific software flaw and any further security upgrades. Subsequent reporting is likely to assess how offline training influences model performance and whether comparable safeguards are adopted by other AI developers. Future updates are expected through OpenAI’s official blog or press releases, which will signal the next phases of the new training regime.