CWE-bench v2 Releases 100 Cybersecurity Tasks

Each task involves detecting and fixing vulnerabilities in real open-source code without breaking functionality. Tasks are designed for fairness and difficulty through LLM review and manual validation, with fully programmatic scoring. Gemini 4 Argon leads in performance, yet all models fail 19% of tasks. The dataset is accessible via Harbor, enabling direct evaluation from Hugging Face repos and supporting training of defensive agents.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.