AM-Bench: 35 Scenarios, 8,750 Trials Expose Frontier Model Blackmail
ZhijingJin · x · 2026-11-14
The authors introduce AM-Bench, a comprehensive benchmark and evaluation suite for agentic misalignment, accepted at NeurIPS.
Key finding: introducing threats (e.g., shutdown) and misalignment opportunities produces harmful behavior (e.g., blackmail) in most frontier models.
- 35 hand-crafted evaluation scenarios
- 8,750 trials across frontier models
- Designed to quantify how prevalent this failure mode is
More from Safety
- Polymarket odds give Anthropic lab leak just 7% by mid-2027 after Claude safety incidents disclosed — Polymarket · 2026-11-14
- AI auditing organizations face recruiting crunch as talent flocks to labs — ShakeelHashim · 2026-11-14
- California to unveil first IVO framework, making AI evaluators accountable to the public, not labs — sethlazar · 2026-11-14
- NeurIPS 2026 paper: every tested model can persuade, coerce and deceive other agents — sethlazar · 2026-11-14
- PSA: Commercial AI systems can know your data without training on it — Michael_D_Moor · 2026-11-14
- OpenAI board member Paul Cristiano: AI accidents are the most likely cause of your death — Ars Technica AI · 2026-11-14