AI coding agents rewrote ten classic Linux utility programs from scratch, and a fuzz-testing study found the results held up as well as - often better than - the human-maintained originals.
The paper, "A Study of the Reliability of Agentic AI-Generated Programs" (arXiv:2609.18298, https://arxiv.org/abs/2609.18298), used a standard agentic AI workflow to regenerate ten well-known, release-quality Linux utilities, then tested both the AI versions and the latest human-written releases with black-box generational fuzzing and AFL++'s coverage-guided testing. The AI-generated programs failed less often overall. They were notably less prone to memory errors like buffer overflows, the classic C bug that fuzzers are built to catch. But they hung more often, falling into infinite loops that the human code mostly avoided.
That trade-off is the real story here. It echoes a debate developers already have about memory-safe languages like Rust: close off one class of bug and a different, harder-to-fuzz class can take its place. The authors are also direct that none of this happens automatically - reliability tracked closely with how carefully the prompts and human supervision were handled, which means the workflow itself becomes a specification worth maintaining.
Ten utilities is a narrow sample, and this is a preprint, not a peer-reviewed result. Treat it as a data point that AI code can hold up under adversarial testing, not proof that agentic coding is ready to run unsupervised.