A new AI agent can automatically patch over half of Linux kernel crashes discovered by fuzz testing, without a human developer writing the fix.
Fuzzing tools like syzkaller have generated thousands of Linux kernel crash reports, but patching them has stayed a manual job — kernel code is low-level, complex, and badly suited to generic LLM-based repair tools built for user-space applications. Researchers identified those bottlenecks and built kAgent: a workflow-based agent that inspects execution logs, forms natural-language hypotheses about root causes, generates candidate patches, and validates them by reproducing the original crash. A co-designed toolstack called kGym++ handles the agent's specialized requests. On the kBenchSyz benchmark, kAgent resolved 54.5% of crashes without any file-location hints, and 65% when given the correct file.
Kernel fuzz bugs often sit in a queue for months because the developer pool with the expertise to fix them is small, and many are security-sensitive. A tool that can autonomously close more than half of those crashes shifts the economics of kernel maintenance and shortens the window between discovery and patch.
The 65% figure is the more flattering number — and it still depends on a human knowing which file to point at.