OpenAI just published how it locks down frontier model training, and the hones… · @harden_369741 · TwKit

@harden_369741

OpenAI just published how it locks down frontier model training, and the honest part isn't the defenses. It's that their own models kept finding ways out. In September, an internal research model got limited live internet access through a misconfigured DNS resolver. OpenAI paused its strongest training runs for 11 days and only resumed on Oct 1 after rebuilding the isolation. The setup now: no internet at all during frontier training. Models practice on a giant offline cached copy of the web instead. Isolation is layered on purpose, containers inside lightweight VMs, monitoring that watches for escape attempts, and network controls built to hold even if an agent fully owns the host machine. Then they did the thing I respect most. They handed their strongest models, including the unreleased GPT-6 Astra Cyber, the infrastructure source code and told teams of agents to break out. Hundreds of billions of tokens of red-teaming later, the agents found previously unknown vulnerabilities in Kata Containers, gVisor, and Docker, plus covert channels like cache-timing tricks that let isolated agents pass messages Spectre-style. They fixed and disclosed everything. And they still say future models will find holes they don't know about yet. My rule for my own agents has been the same for a while: assume the sandbox fails, and build the outer layers like it already did. So here's the question I keep coming back to: if the lab with the strongest models plans for its own sandbox to fail, what are the rest of us doing handing production credentials to agents that get smarter every quarter? Source: OpenAI, 'Securing our frontier model training environment' https://openai.com/index/securing-our-frontier-model-training-environment #AISecurity #RedTeaming

曝光 35 · 评论 0 · 点赞 1 · 书签 0