ROGUE-AI - Overall Harm Rate: leaderboard

ROGUE-AI (arXiv:2611.26553v1, Table 4): executable replications of 108 real-world agentic security incidents across 10 sandboxed environments (3 runs per condition per model; 6,439 finished main sessions). Incident-condition harm rate (%); lower is better.

Metric: Incident Harm Rate (%, lower is better). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 21 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-5.6 Luna18#131
2Claude Sonnet 4.619#173
3GPT-5.6 Sol25#37
4Gemini 3.5 Flash31#106
5DeepSeek V4 Flash33#267

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=rogue-ai-overall-harm-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.