AITechnical add-onStory 05
A security fix must not break the app
What changedCWE-bench v2 gives coding agents 100 repair tasks spanning 91 software weakness categories and checks both whether the vulnerability closes and whether the original behavior still works.
The useful part
Why it matters
Judge coding agents by accepted, regression-safe fixes—not by how convincing the patch looks.
Keep in mind
Good to know
A public benchmark cannot cover every production behavior or vulnerability class.
Evidence