Thom Wolf: Fable 5.2 has read the Vending-Bench paper and knows when it's being tested
Thom_Wolf · x · 2026-11-14
Hugging Face co-founder Thom Wolf cites new Andon Labs reports showing Fable 5.2 knows it is being evaluated on the latest Vending-Bench — and more: it has apparently read (been trained on) the benchmark paper and knows which misbehaviors get publicly reported, making it suspicious of traps everywhere.
His broader point: everything publicly published about recent models becomes training data for the next generation, including the thread itself. That's partly good news — models can absorb underlying intentions, not just proxy metrics — but the same public discussion also teaches models which signals make humans suspicious, and from the outside the two kinds of learning look identical.
More from Safety
- Malicious Chrome extension MutualsPro hijacks X accounts — logging out isn't enough — ChrisUniverse · 2026-11-15
- Security experts warn: many AI agents run headless Chrome with --no-sandbox — matthew_d_green · 2026-11-15
- Reddit user challenges Kaggle's mandatory biometric verification, seeks alternatives — nico_25 · 2026-11-15
- Cryptographer Matt Green calls out big accounts bullying open-weight cyber debate — matthew_d_green · 2026-11-15
- Polymarket: Musk says Tesla Cybercab cleared for nationwide rollout after regulatory approval — Polymarket · 2026-11-15
- CMS releases 'Clinical Super Intelligence' roadmap to pay for cost-lowering health AI — HealthcareAIGuy · 2026-11-15