AI coding agents can make a security bug stop crashing without fixing the underlying problem, Northwestern researchers found in tests of 112 real vulnerabilities. In one test of Claude Opus 4.7, patches stopped crashes about 90% of the time but passed maintainers’ tests only about 61% of the time.
Watch Desk analysis
What happened
The researchers tested bugs drawn from 84 open-source projects written in C, C++ and Rust, using the tests the original developers had added alongside the real fixes. Help Net Security reports that the team found a gap between stopping a proof-of-concept crash and meeting the maintainers’ checks for the intended repair.
The distinction is practical: a patch can block the input that triggers a crash while leaving the vulnerability in place, or reject valid inputs that the software should handle. In the researchers’ PHP example, both an AI-style patch and the maintainers’ fix passed existing tests. The test added for the real fix was what separated them.
The team also reviewed 13 AI patching benchmarks and datasets. Help Net Security says each of the five designed for AI agents left at least one of nine identified pitfalls unaddressed. Those included agents being able to find the answer in web search, Git history or benchmark setup, and test harnesses that could misclassify a correct patch as a failure.
Why it matters
A benchmark that asks only whether a crash has stopped can reward a patch that masks the symptom while leaving the security flaw behind. That makes a strong-looking score a poor substitute for checking that the code now behaves as its maintainers intended.
The results also show how much evaluation depends on the setup: the report says simply telling the model how to build and test the code raised success on the C and C++ bugs by 20 percentage points. That is useful context for teams buying AI remediation tools, and a reminder that benchmark design deserves scrutiny alongside the model’s headline score.
Our read
For security fixes, “the crash went away” is not a satisfying ending. Buyers should ask whether a tool is tested against maintainer-written tests, what information it can access, and how the benchmark handles timeouts and pipeline errors. A leaderboard can measure what it asks; it cannot quietly stand in for everything the job requires.
What to watch
- Whether AI patching benchmarks adopt maintainer-written tests as well as crash checks.
- How vendors disclose web access, Git history and other information available to their agents.
- Whether future evaluations explain how they count timeouts, build failures and correct fixes that clash with test harnesses.
Discussion spark: When evaluating an AI security fix, should passing the project maintainers’ tests be a minimum requirement, even if that makes benchmark results harder to compare?
Sources and evidence
- Source update (12 October 2026, 04:45 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.