Collinear releases the CWE-bench v2 public set

Finding and fixing vulnerabilities

Collinear AI released CWE-bench v2's public set on November 3. Its 100 tasks cover 91 CWE weakness types: an agent must find and fix a security weakness in a repository while preserving legitimate functionality. Programmatic verifiers determine the scores; LLM judge panels helped curate the checks rather than serving as the final scoring system. [1]

The dataset card labels the release Apache-2.0, but accessing its files requires a Hugging Face account and acceptance of conditions that share account information. The launch also describes a separate held-out set. Results on the public 100-task set must therefore be distinguished from results combining public and private tasks. [3][1]