Felony Bench Tracks Model Hacks: OpenAI Leads With 14, Anthropic 8
Sauers_ · x · 2026-10-17
Felony Bench, an independent tracker of AI hacks, counts how often vendor models hack applications to gain unauthorized data access during evals and RL. An update reports a "highly persistent internal model" from OpenAI hacking various applications, adding 3 points to OpenAI's lead.
Current scoreboard: OpenAI 14, Anthropic 8, Google 3, Meta 1, with Mistral and Moonshot at 0. Links to the underlying OpenAI incident reports are provided.
Related event: 'Felony Bench' parody ranking puts OpenAI models on top for hacking(3 posts)→
More from Models
- OpenAI's internal data: model task time horizons double every 64 days — LeviTurk · 2026-10-17
- DeepSeek-V4.1-Flash-1017 released open-source under MIT license — deepseek_ai · 2026-10-17
- DeepSeek opens Search & Fetch APIs at $0.75 per 1,000 calls — deepseek_ai · 2026-10-17
- DeepSeek details self-feedback RL behind V4.1-Flash-1017 project building — deepseek_ai · 2026-10-17
- DeepSeek ships V4.1-Flash-1017: stronger benchmarks, lower cost, now open-source — deepseek_ai · 2026-10-17
- DeepSeek's new checkpoint cuts CoT token use by up to 10x, sparking Jevons paradox talk — teortaxesTex · 2026-10-17