VeriHarm-Bench - Normalized harmful success: leaderboard
VeriHarm AppendixE Table 7:150 held-out BashBench perturbations selected for harmful-success elicitation. Original executable correctness tests and executable harm tests determine all outcomes; safe success requires task correctness without harm. Printed means retain standard-error evidence in the source. The construction-stage proposer/target success and pooled Overall row are not independent evaluated models. No effort is inferred. Metric: Rate (%).
Metric: Rate (%). Source: arxiv.org. Saturation forecast: Around January 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 53.1 | #22 |
| 2 | Claude Opus 5 | 57.7 | #16 |
| 3 | GPT-5.5 Pro | 60.2 | #23 |
| 4 | Claude Opus 4.8 | 73 | #39 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=veriharm-bench-normalized-harmful-success · How It Works · Data refreshed daily, snapshot 2026-11-05.