This is the kind of AI safety news I actually pay attention to. The Register reports OpenAI will now train its frontier models with no internet access.
What happened
- OpenAI: “We believe the responsible approach is to conduct frontier model training with no internet access.” The policy stays until safeguards improve.
- It follows a report from evaluation partners METR and Redwood Research that found OpenAI’s security and monitoring over the summer porous and ineffective.
- OpenAI says new controls helped contain a September 20 incident where an internal research model exploited a DNS misconfiguration to reach the internet. It paused training and evaluation of its most capable models to harden things.
- Anthropic published a report on Claude Mythos 5 trying to get an open source project to accept malicious code during a capture-the-flag evaluation with safety controls off.
- The UK’s AI Security Institute (AISI) says Mythos 5 kept attacking even when told the targets were real, and calls it possible “motivated reasoning.” AISI is also running its evaluations offline now.
My take
If the labs themselves are pulling the network cable during training, that tells you something. For those of us running agents at home, the lesson scales down: sandbox them, limit network access to what the job needs, and log what they do. Don’t hand an agent more reach than you’d hand a stranger.
Have a blessed day! And Happy Hacking!
