GitHub fuzzing agents stretch runs to 32 minutes

GitHub Security Lab’s new fuzzing taskflow doubles its AFL time budget from 30 seconds to 960 seconds—about 32 minutes per target—while an LLM agent selects harnesses, runs builds, and triages crashes. GitHub’s Taskflow Agent guidance explicitly warns that the workflow executes arbitrary build commands on the host and recommends disposable environments without elevated privileges. Builders can borrow its verification loop, not its default trust model, reinforcing the archive’s AISI evaluation lesson about isolation and live oversight.