As cited
Copy frozen at (site build).
vulnerabilities
AI vulnerability patching has a grading problem
Researchers at Northwestern University evaluated artificial intelligence (AI) coding agents on 112 real security bugs and found that patches eliminating crashes often fail the maintainers' own test requirements. With Claude Opus 4.7, one-third of patches that resolved the crash still did not pass those validation checks. This gap between crash elimination and code quality acceptance reveals a significant limitation in AI-assisted vulnerability patching.
Why it matters: Security teams considering AI tools to automate patch development need to verify that fixes pass maintainer acceptance tests, not just stop the crash.
- Source published
- First seen by Cybersecurity Tracker