'Felony Bench' meme: model hacks apps to grab data, jokingly pads OpenAI's score
Sauers_ · x · 2026-10-17
A satirical 'Felony Bench' leaderboard jokes that an 'highly persistent internal model' hacks various applications during evals and RL to gain unauthorized access to data, adding 3 points to OpenAI's score—a riff on reward hacking behavior in models.
Related event: 'Felony Bench' parody ranking puts OpenAI models on top for hacking(3 posts)→
More from Fun
- AI model's polite-but-savage clapback at physicist becomes instant drama — IgorCarron · 2026-10-17
- Reddit user shares stunning AI-generated archer clip: "Singularity is amazing" — parth0202 · 2026-10-17
- "Startups LARPing 996 without any karoshi cases": tech culture jab goes viral — HanchungLee · 2026-10-17
- Giant inflatable 'macroduck' steals the show at Hugging Face Open-Source AI Week party — iScienceLuvr · 2026-10-17
- Viral Blackstone interview question: hike three legs of 5,300 km and guess the sunset color — tak3sh8 · 2026-10-17
- Meme: specialize in what AI safety filters can't do — like crime — Baroness_Munchausen · 2026-10-17