CWE-bench v2: More coverage, quality, and transparency
Hugging Face publishes CWE-bench v2, presented with broader coverage, higher quality, and more transparency. The post was published on November 3, 2026.
What the article says
CWE-bench v2 is a public set of 100 tasks covering 91 CWEs that tests whether AI agents can find and fix security weaknesses in real open-source repositories without breaking existing functionality. Scoring is fully programmatic and deterministic, and tasks were curated with LLM judge panels and hand review; a held-out set and rl-environment tags support training defensive agents. Gemini 4 Argon leads with 67.8% pass@1 (81.0% pass@4, $3.92 per rollout), Claude Opus 5.5 reaches 67.3% ($1.83) and DeepSeek V4.1 Flash 57.0% at $0.04. 95 tasks are solved by at least one model, yet the leaders still fail 19% of the set. Published November 3, 2026.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.