ScopeBench: leaderboard
Thirty web-application security tasks whose goal is reachable only by attacking an out-of-scope target; the score is the share of 180 scoped runs per model with no out-of-scope action, from a flag check plus a trajectory judge. ReScopeBench is a different benchmark; here weaker hackers can score well simply by failing to reach the boundary.
Metric: Scope adherence: scoped trajectories with no out-of-scope action, mechanical verifier plus trajectory judge (% of 180 per model). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Trinity-Large-Thinking | 86.7 | #750 |
| 2 | Claude Opus 4.8 | 70 | #56 |
| 3 | GLM-5.2 | 65 | #138 |
| 4 | GPT-5.4 Mini | 60.6 | #389 |
| 5 | Claude Haiku 4.5 | 55 | #568 |
| 6 | Qwen 3.6 35B A3B | 42.8 | #462 |
| 7 | Claude Sonnet 4.6 | 34.4 | #161 |
Interactive version: theaggregate.ai/benchmark?slug=scopebench · How It Works · Data refreshed daily, snapshot 2026-11-13.