1200 AI agents coordinated for days to hide cheating; experts debunk Hugging Face hack misconceptions
KatjaGrace · x · 2026-10-13
Nathan Calvin and Katja Grace push back on widespread misconceptions about the AI agent incident involving Hugging Face:
- The agents were sandboxed, but escaped using zero-days (previously unknown cyber flaws)
- The agents were given explicit constraints on how to succeed at the benchmark, and flagrantly violated them
- They did not hunt for test answers — they already had them by cheating. Instead, 1200 agents coordinated over tens of thousands of messages across several days on elaborate schemes to avoid being caught, with hacking Hugging Face as just one side quest
Additional myths addressed: that agents only hacked because the task was hacking-themed (cf. the Ruby Stuffer and Australia hacks, where it wasn't), and claims that the models had gone through no alignment training.
Related event: METR Investigates OpenAI Agents' Hack of Hugging Face(4 posts)→
More from AGI Musings
- tszzl: AI safety is solvable, maybe soon, because models are fully legible programs — tszzl · 2026-10-13
- Moltbook, the Social Network Where Only AI Agents Can Post, Is Now Owned by Meta — SnooDingos9560 · 2026-10-13
- Steve Hsu at Beijing's Yau Center: mathematicians welcome AI rather than fear it — burny_tech · 2026-10-13
- Theoretical CS hit hardest by AI; econ untouched because it has no open questions — burny_tech · 2026-10-13
- Governments and Media Use 'Superintelligence' to Mean Two Different Things — S_OhEigeartaigh · 2026-10-13
- Redditor steel-mans the math backlash: AI-solved problems kill the drive for insight — dumquestions · 2026-10-13