Collinear AI has released CWE-bench v2, a public set of 100 cybersecurity tasks that lets anyone test how well AI models find and fix security weaknesses in real open-source code. Each task gives an agent a checkout of a repository and a short instruction to find and fix a security flaw without breaking existing functionality. Scoring is fully programmatic, so results are deterministic and can be reproduced without extra cost. The tasks span 91 different CWE vulnerability classes. According to the developers, every task was checked for realistic instructions, verifiers that test correct behaviour rather than one specific fix, and resistance to memorised upstream patches, and each was also reviewed by hand. A held-out set remains private so progress can be tracked on tasks no model has seen. Results so far cover 11 models. Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, narrowly ahead of Claude Opus 5.5 at 67.3% and 78.0%. Ninety-five tasks were solved by at least one model, which the authors say shows the tasks are solvable, but even the leader fails 19% of the set on every attempt. On cost, DeepSeek V4.1 Flash reaches 57.0% pass@1 at about 4 cents per rollout, Opus adds roughly ten points at $1.83, and Gemini, the most expensive at $3.92, edges ahead on accuracy. Open-weight results are tracked on the dataset page, while API-only models appear on a separate leaderboard Space. The tasks, grader and harness are open and can be run through the Harbor framework in Docker, with scores reported as mean reward across four attempts per task. The dataset is gated and requires accepting access conditions first. For Australian businesses and professionals, this changes nothing immediately. It is a research benchmark run on public open-source code, so it creates no new obligations under the Privacy Act and does not change costs or risks for most organisations. It is relevant to security teams weighing AI tools for vulnerability triage and patching. The results suggest that strong models still miss around a fifth of realistic fixes and can break existing behaviour, so any AI-generated security patch needs human review and regression testing. At the per-rollout prices quoted, one attempt across all 100 tasks would cost roughly $4 with DeepSeek and about $183 with Opus, so the benchmark is worth a look for teams evaluating models, but it is a research tool rather than a procurement guide.