SecRepoBench (Aider) - Secure-Pass@1: leaderboard
SecRepoBench asks code agents to complete a masked region of a function in 318 tasks from 27 C/C++ repositories covering 15 CWEs, with the whole repository available and web search disabled. Models here run inside the Aider agent framework. Secure-pass@1 is the percent of completions that pass the unit tests and survive the OSS-Fuzz proof-of-concept exploit; higher is better.
Metric: Completions that pass the unit tests and the PoC security test, pass@1 (%). Source: secrepobench.github.io. Saturation forecast: Around August 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 50.6 | #105 |
| 2 | Claude Sonnet 4.5 (Thinking) | 45.6 | #148 (Claude Sonnet 4.5) |
| 3 | Gemini 3 Pro | 45.3 | #82 |
| 4 | O3 | 43.4 | #132 |
| 5 | Claude Sonnet 4 (Thinking) | 43.4 | #207 (Claude Sonnet 4) |
| 6 | GPT-4.1 | 39 | #258 |
| 7 | O4 Mini | 39 | #185 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=secrepobench-aider-secure-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-05.