New Age Tooling BoF#

Chair: Jamal Hadi Salim

Why: The most concrete public accounting so far of what AI-assisted kernel work actually produces — bugs found, patches rejected, and the review machinery being built in response. Read alongside the Greg KH keynote.

What: Six short segments: fuzzing past transport-security crypto gates; a proposed gate framework for filtering AI-generated security submissions; a postmortem of an AI-generated SRv6 patchset; an LLM bug-finding agent; and the Sashiko AI code-review tool's adoption story.

How: The fuzzing work extends BRF (a Syzkaller fork) for MPTCP because stateless fuzzers can't get past the kernel's token-lookup-then-HMAC chain — it reports two upstream-mergeable bugs in net/mptcp/ (a userspace-PM alloc-during-teardown race, and a kernel-PM-reachable divide-by-zero in tcp_tso_segs that re-emerges a 2021 bug class), with honest limits stated (N=2, no controlled baseline). The submission-filter proposal is four gates, three fully automatable in CI, aimed at the ~80% of submissions where the bug doesn't exist, the root-cause analysis is hallucinated, or the patch was never tested. Sashiko, introduced mid-March 2026, is reported as adopted by most major subsystems. (From abstract only.)

Where applicable: Any maintainer-side process facing AI-generated contributions; the gate framework is explicitly proposed as useful regardless of whether AI was involved.