Microsoft's FORGE Lab reported this week that MDASH, its AI scanning system, produced 59 accepted bug reports against the upstream Linux kernel, including one flaw that had survived at least 21 years. The number sounds large until you compare it with the queue it joins: through November 2, the Linux CVE team averaged 27 entries a day in 2026.

Finding bugs stopped being the hard part some time ago. Reviewing and fixing them is the part that has not scaled.

Every bug is a security bug eventually

Linux gives a bug report unusual reach because any defect can become a security defect. "Bugs will happen, and anything can be a security bug if somebody is clever enough to just figure out how to abuse it," Linus Torvalds said in 2024. The same code runs on phones, refrigerators and supercomputers, so a fault that crashes one device may never surface on another, and triage takes context that only a maintainer has.

Windows shows the same shape. Microsoft's latest Patch Tuesday listed 726 CVEs. Most of the reports arriving at both projects now come from AI programmers working at machine speed.

FORGE Lab's own summary is the cleanest statement of the problem: "Finding bugs and fixing them are drastically out of balance." Turning its scanner's output into 59 usable reports took more than four person-months, because human effort was the bottleneck in the loop.

The reviewers are leaving

On November 10, Lorenzo Stoakes stepped back from reviewing code. Stoakes is an Arm engineer who has contributed to the kernel since 2014 and reviewed patches for the memory-management core, the misc tree, reclaim and the Rust-related sections. In a mailing list post he wrote that given "the current throughput in the kernel resulting from AI-generated series," he no longer had the bandwidth to review anything outside the areas he maintains.

He was blunter about why. "Of course the AI wave has been horrible, and is getting significantly worse," he wrote. "At first it was once in a while that I'd see slop. Then it was at least once a week. Then every few days. Nowadays it's pretty much half of what I see, if not more."

Greg Kroah-Hartman, who maintains the stable tree, put the staffing problem to an audience at Open Source Summit Europe. "We don't have enough people to review them, the onslaught of the LLM-generated patches," he said. "We need more reviewers. We don't have enough. All of the maintainers are burnt out, and we have been for ten years."

Two reasons for cautious optimism

Torvalds wrote on the kernel mailing list that much of the new code in Linux 7.4 rc2 "was driven by various AI fixes." On the BPF side, he said, fixes are arriving faster than new reports, which he called a lovely sign if it holds. "I think most other areas haven't reached that point yet, but it's at least a sign that we may be starting to see the tide starting to turn."

Kroah-Hartman is not ready to agree. "We are getting closer, yes, but not quite there yet," he said, and declined to guess when human-approved AI fixes would overtake AI bug reports. Daniel Stenberg, who created curl and has been loud about low-quality submissions, told ZDNET the project manages the flood "pretty well so far," with swings that depend on who sends a batch of findings that week.

The security record for November argues for patience. A serious TCP networking hole was disclosed, along with an OpenSSH flaw in Alpine Linux that patched a bypass of the login check, a near miss at Alpine's GitLab server, and PHP2Shell, which many operators still have not fixed. Each of those needed a human to write and review the repair.

The measurable question for the next quarter is not how many bugs a scanner can find. It is whether the kernel can add reviewers at the rate Stoakes just subtracted one. Torvalds has also been pushing in the other direction, dropping the AI disclosure tag from patches in October while posting an LLM-assisted patch of his own, which suggests the project intends to absorb the tooling rather than filter it out.

Source: ZDNET