COMMITGUARD catches commit-induced bugs, cutting leads from 518 to 7

Code reviews and existing tests often miss subtle memory-safety issues like boundary, lifetime, or initialization errors lurking in a new commit. Whole-program fuzzing can catch them but is too expensive to run on every change. COMMITGUARD, a new tool from Aniruddhan Murali, Noble Saji Mathews, Mahmoud Alfadel, and Meiyappan Nagappan, tackles this by focusing on the exact functions a commit modifies. It extracts compilable code slices from both the pre-commit and post-commit versions of each changed function, fuzzes the two slices independently, and compares sanitizer reports. Bugs that appear only in the post-commit slice are flagged as commit-induced candidates, using the pre-commit version as a behavioral baseline.

Evaluated on 300 commits from OpenSSL, libpcap, and leptonica, COMMITGUARD initially produced 518 sanitizer reports. After differential comparison, that shrank to just 7 candidates requiring manual triage. Developers confirmed 5 as real bugs—since fixed by the project teams—and only 2 were false positives. The tool averages 32.4 minutes per commit and reaches 75.36% coverage of modified functions, making it practical for continuous integration. By filtering out pre-existing issues, COMMITGUARD turns noisy fuzzing output into a short, actionable list of regressions, promising to catch critical memory bugs before they ship.