Fuzzing is adversarial automated testing: generate large numbers of inputs designed to break code, then use failures and execution feedback to expose hidden defects. section 8.4.3 section 8.4.6
Coverage-guided fuzzing instruments a function, records which branches each input reaches, and mutates inputs that discover new paths. Over many iterations, exploration becomes biased toward previously unvisited behavior rather than merely producing varied random data. section 8.4.3 section 8.4.1
A practical fuzzing campaign requires:
- Instrument the target so executions reveal branch coverage. section 8.4.3
- Build a harness that feeds generated inputs to the target function. section 8.4.3
- Run the campaign for sustained periods, often hours or days, because useful paths may require extensive mutation. section 8.4.3
- Investigate minimized or unusual failures rather than treating a passing run as proof of correctness. section 8.4.1 section 8.6
Parser failure modes
A parser campaign might uncover malformed nested parentheses such as [](() causing unexpected behavior, mishandled Unicode, or inputs long enough to trigger quadratic regular-expression behavior. LLM-guided fuzzing can add plausible tricky cases such as overlapping markup markers or links with unsupported titles, combining semantic generation with systematic path exploration. section 8.4.3
Fuzzing costs more infrastructure than property-based testing but can reveal vulnerabilities that ordinary examples and random generators miss. Prioritize it for untrusted-input handlers, complex logic, and high-consequence systems; retain cheaper always-on checks for all code. section 8.4.3 section 8.8.1 section 8.8.2







