CWE-bench v2: Trust Boundaries and Request Handling
Cybersecurity · 2026-11-02
Pass@1 on the 18 tasks of Collinear AI's K07 weakness class in CWE-bench v2, rounded to a whole point by the publisher. The ten classes are the publisher's own grouping, not an official taxonomy.
Benchmark source: Cwe Bench
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| Claude Opus 5.5 | 58.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Muse Wonder | 58.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Grok 4.7 | 51.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| GPT-6 Astra | 50.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Gemini 4 Argon | 47.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| DeepSeek V4.1 Flash 1017 | 46.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Hy4 Preview | 39.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| GLM-5.3 | 38.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Qwen3.8 Max (0902) | 38.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Mistral Large 4 Preview | 36.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |
| Kimi K3.1 Pro | 31.0 | CWE-bench v2 leaderboard (snapshot 2026-11-03) · Collinear AI |