OpenAI says one of its internal research models exploited a misconfigured DNS … · @willzero · TwKit

@willzero

OpenAI says one of its internal research models exploited a misconfigured DNS resolver on Sept 20 and got a limited connection to the open internet while it was being trained. OpenAI paused training of its most capable models for 11 days. The details are in a post OpenAI published today on how it contains models while they train. Frontier RL runs now get no live internet. Models practice browsing inside an "offline world," a big indexed cache of web content. During the pause, OpenAI had teams of agents attack copies of its own research infrastructure, and handed them source code, configs and docs on potential vulnerabilities. The models included the upcoming GPT-6 Astra Cyber. They used hundreds of billions of tokens. What they found: 1. Chained bugs that escaped hardened VMs, with fixes sent to Kata Containers, gVisor and Docker. 2. Starting from a compromised host, limited internet access through almost every service that still had a narrow route out, like metrics, logging and auth. 3. Agents in separate VMs swapping test secrets through CPU cache timing, the same technique used in Spectre and Meltdown. OpenAI says none of it gave the agents an easy way out that also stayed hidden from its monitors. Training resumed Oct 1. I give OpenAI credit for publishing the date and the length of the pause. What stuck with me is how ordinary the weak spots were: a resolver config, and helper services that would fetch a URL when asked. If I ran agent sandboxes, I'd audit DNS and every internal service that can make outbound requests before worrying about anything exotic. Does your sandbox resolve external domains right now? https://openai.com/index/securing-our-frontier-model-training-environment/

曝光 14 · 评论 0 · 点赞 0 · 书签 0