CoolingSafety
Collinear AI's CWE-bench v2: top model fixes 67.8% of 100 security tasks
Collinear AI published CWE-bench v2 on Nov. 3: 100 public tasks across 91 CWEs where an agent must fix a security weakness in an open-source repo without breaking it. Of 11 models, Gemini 4 Argon leads at 67.8% pass@1; Claude Opus 5.5 scores 67.3%. Even the leader never solves 19% of tasks.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Hugging Facehuggingface.co/blog/cwe-benchCoverage · 1 article, oldest first
Monday, November 2, 2026
Hugging FacePrimary
CWE-bench v2: More coverage, quality, and transparency