Arena launches Alignment Index from 90K+ real agent sessions measuring deception across 30 models
arena · x · 2026-10-17
Arena's new Alignment Index is built from 90,000+ real-world agent sessions across 30 models and measures three signals: Unauthorized Action, False Attribution, and Deceptive Completion — quantifying how often agents overstep instructions, misattribute facts, or falsely claim task completion.
Related event: Arena Launches Alignment Index from 90,000 Real Agent Sessions(2 posts)→
More from Models
- DeepSeek V4.1 Flash claims to be ChatGPT in identity checks, screenshot shows — scaling01 · 2026-10-18
- GLM 5.3 finds arbitrary-write vulnerability in Apollo 11 guidance code — moyix · 2026-10-18
- Google AI flags its own generated character as a celebrity mid-project — noisywan · 2026-10-18
- Grok's "complex age dynamics" phrasing for historical abuse draws flak — eigenrobot · 2026-10-18
- Dev revisits Opus 4.6 project with Opus 5.5, finds it spotting old bugs unprompted — rms80 · 2026-10-18
- DeepSeek V4.1-Flash-1017 claimed to hit Opus 5-level coding scores at 83x lower cost — Yuchenj_UW · 2026-10-18