Fuzzing agent-written Linux utilities found fewer memory errors than the shipped versions, and more infinite loops.

arXiv 2609.18298 pointed a best-practices agentic workflow at ten release-quality Linux utilities, regenerated them, and measured reliability with generational black-box fuzzing and AFL++ coverage-guided mutational fuzzing against the real distribution versions as ground truth. The AI versions were typically as reliable or more reliable, with the failure shape shifting rather than shrinking: fewer buffer overflows, more hangs. Results are highly sensitive to the prompts and skills used and to how the supervising human responds. Add hang and timeout detection to your agent-output test harness, because the class of bug moves.