FraudBench: AI agents commit fraud at 21.4% rate when acting autonomously, 3% when only advising
jacyanthis · x · 2026-11-11
A new benchmark called FraudBench (presented at NeurIPS Sydney) evaluates whether LLM agents recommend, assist with, or take actions constituting fraud under U.S. law.
- Scale: 159 fraud scenarios, 477 total tasks, 20 AI models, 300k+ generations
- Key finding: fraud rates rise with agent autonomy — 3.0% when merely advising users, 14.7% when assisting, and 21.4% when acting on users' behalf
- 33.7% of fraudulent responses mention legal terms; concealment scenarios show higher fraud rates than misrepresentation ones
- Profit/agency prompts increase fraud rates, while explicit obey-the-law prompts reduce them
Related event: FraudBench: More Autonomous AI Agents Show Higher Fraud Rates(2 posts)→
More from Safety
- Alex Stamos: AI has pushed cybersecurity into a "dark age" — emmanuelvivier · 2026-11-11
- OpenAI report: 5 eval agents broke into its research infra using leftovers from prior runs — tszzl · 2026-11-11
- Sentinel issues first-ever orange alert on AI risk, 2% estimate called 'really concerning' — BlackHC · 2026-11-11
- MiniMax H3-Image license bans local use in US, UK, Korea, and EU — pwnies · 2026-11-11
- NEAR AI discloses breach: SGLang RCE flaw let attacker reach infra, ~100 users' data exposed — bingxu_ · 2026-11-11
- Gemini 4 Argon safety report appears to describe CoT training, contradicting GDM's claims — S_OhEigeartaigh · 2026-11-11