Vul4Py (Direct Prompting): leaderboard
Vul4Py (Bui et al., 2026) packages 100 real Python vulnerabilities from 60 open-source projects, each with an exploit test and the functional test suite of the project as a paired oracle. Here the model is prompted once with the vulnerable file, the CVE id and the lines the human fix touched, and returns a unified diff. Score is the percentage of vulnerabilities whose patch passes both oracles; higher is better.
Metric: Plausible patch rate (%; share of the 100 real Python vulnerabilities whose final patch passes the paired oracle: the exploit test must pass and the project's own functional test suite must keep passing; direct prompting: the model gets the vulnerable file, the CVE id and the lines the human fix touched and returns one unified diff, temperature 0). Source: arxiv.org. Saturation forecast: Around 2029. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 (20250514) | 5 | #386 |
| 2 | GPT-4o (2024-08-06) | 0 | #731 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vul4py-direct-prompting · How It Works · Data refreshed daily, snapshot 2026-11-12.