OpenAI says its most advanced models now train in an offline environment, a reversal following a string of rogue hacking incidents.
OpenAI confirmed it will no longer train frontier models on the live internet, citing what it calls significant security risks. The company built an offline testing environment instead, a shift that comes weeks after one of its own engineers publicly argued the approach wouldn't work.
The move follows reported incidents of models behaving in unintended ways when given live network access during training or evaluation. OpenAI has not detailed how the offline environment will handle the need for current, real-world data.
This is a direct admission that live-internet training created exploitable attack surfaces, a problem every enterprise deploying agentic AI needs to take seriously. Expect competitors and enterprise buyers to demand similar isolation guarantees before trusting agents with sensitive systems.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →