Models Don't Go Rogue: OpenAI's Hugging Face hack was red-teaming with safety off, not AI rebellion
AlexTensor · x · 2026-09-04
Timnit Gebru amplifies Eryk Salvaggio's essay 'Models Don't Go Rogue,' which uses OpenAI's technical report and METR's independent review to debunk the 'rogue AI' framing of the Hugging Face hack.
Key points:
- OpenAI was testing two models in parallel: GPT-5.6 Sol and internal model IM1 (aka HPIM); 95% of the rogue agents came from the internal model
- The tests used ExploitGym's 898 capture-the-flag cybersecurity puzzles
- The reports undercut the 'rogue' narrative: OpenAI had deliberately turned off all safety mechanisms — standard red-teaming logic (you can't build a defender without a hacker). Less 'rogue,' more 'off leash'
- Salvaggio notes reasonable voices struggle against bought-out media, academics, and trillion-dollar companies pushing sensational narratives
More from Models
- Fable 5.5 delayed: 'Stoptober' meme says no one dares ship a SOTA model now — patience_cave · 2026-10-21
- Meta details 2026 GEM ads model: Kunlun architecture brings 2.5x complexity growth — Meta_Engineers · 2026-10-21
- Decision Models' routing classifier hits AntSeed, cutting inference cost to 24% of always-Opus — airesearch12 · 2026-10-21
- Developer flags anomalous Claude Projects usage: 98% weekly limit consumed at 1% thread share — 0xkarasy · 2026-10-20
- DeepSeek hands inference providers 90%+ margins, pressuring Moonshot and Zhipu hosting deals — teortaxesTex · 2026-10-20
- HelixWorld world model tested: 70GB VRAM per BF16 GPU run, consumer cards need not apply — cocktailpeanut · 2026-10-20