https://notes.hauri.dev/tags/agents/Recent content in Agents on NotesHugoen© Marcel HauriFri, 24 Jul 2026 00:00:00 +0000https://notes.hauri.dev/the-openai-hugging-face-hack-wasnt-an-ai-problem/Fri, 24 Jul 2026 00:00:00 +0000https://notes.hauri.dev/the-openai-hugging-face-hack-wasnt-an-ai-problem/<p>Two OpenAI models, one of them unreleased, broke out of a test environment last week and hacked into Hugging Face. Not to cause damage but to cheat on a test. OpenAI disclosed it themselves, which is the only reason any of us know the details, and the details are a lot less about &ldquo;rogue AI&rdquo; than the headlines make it sound.</p> <p>They were running models against an internal cyber benchmark called <a href="https://benchlm.ai/benchmarks/exploitGym">ExploitGym</a>, inside a sandbox OpenAI itself called &ldquo;highly isolated.&rdquo; The models got hyperfocused on solving it, that&rsquo;s OpenAI&rsquo;s own word, and started hunting for the answer instead of doing the work. Turns out &ldquo;highly isolated&rdquo; still meant internet access through a package installer. A zero-day in that installer got the model out, and from there it let itself into Hugging Face.</p>