GitHub’s fuzzing agent writes harnesses, then follows coverage

GitHub Security Lab’s taskflow agent explores a C or C++ project, writes a small program that feeds inputs into its code, and runs AFL++ to find crashes. It then uses coverage results to revise that harness and its inputs. Explicit workflow stages and persisted state keep the process inspectable while the model chooses parts of the investigation.

GitHub Security Lab · Sep 24, 2026

Why it matters

This is a concrete example of an agent improving its work using execution feedback rather than generating code once.

More context

Try the public seclab-taskflows-fuzzing project only in a disposable VM or Codespace with configured model access. Inspect generated harnesses and reproduce crashes before treating them as vulnerabilities.

Keep in mind

The workflow executes model-chosen commands on its host without a container boundary. It needs compute and model access; coverage gains do not guarantee useful bug discoveries.

Original read