Benchmark
AutoPatchBench
Evaluate agents that repair security defects in native code.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
C/C++ vulnerabilities discovered by fuzzing: 142 full-set cases, a 120-case lite set, and a 20-case sample set.
Methodology
Generated patches undergo quality checks, extended fuzz testing, and differential debugging. The reported correct-patch share requires both fuzzing and differential checks to pass.