Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Linux kernel patches under pressure as AI speeds up bug discovery and exploit writing

Linux distributors are shipping fixes for a TCP flaw rated 9.8, while maintainers argue over how much AI is really changing kernel security work. The debate follows a month in which Postgres reverted features and OpenClaw disclosed 23 confirmed vulnerabilities found by an AI-assisted audit.

TechnologyExplainerRachel NwosuPublished: 5 October 20269 min readSources 10
Linux kernel patches under pressure as AI speeds up bug discovery and exploit writing

On 3 November, Red Hat published patches for a Linux kernel TCP remote code execution flaw carrying a 9.8 CVSS score. Ubuntu released its own fixes the same day. The bug, tracked as CVE-2026-98432, is an IPv4 TCP use-after-free. The timeline is short: scanning for the flaw was already being reported in the wild on 2 November.

That is the concrete part. The argument is messier.

Over the past month, the question of how much AI is reshaping kernel and database security work has moved from conference talks into commit logs. Some of the loudest claims are being pushed back by the maintainers who actually reverted the code. The picture is not one story about AI breaking open source. It is a set of separate, smaller stories about disclosure speed, review load and tooling, each with its own evidence and its own gaps.

What actually got reverted in Postgres 19

Several major features were pulled from Postgres 19 over the past month. Tomas Vondra, a PostgreSQL contributor, says the release will now slip by about a month even if everything else goes well. Writing on his blog, Vondra takes issue with a narrative he says has spread through other posts: that AI tools found complex bugs in the patches, and the fixes were too invasive to land this late.

He went through the revert threads one by one. "None of the issues reported earlier seems to have been found by AI," he writes. He adds that the review that finally led to one revert came from Andres and was "clearly a human-written review." Vondra does flag a thread about breaking online checksums with Claude. He notes it came multiple days after Robert's "scary patch" message mentioning the patch in question. His conclusion is that the reverts followed too many post-commit issues, some found late, and a cautious call to reduce risk.

The claim he is arguing against came from Elizabeth Christensen. She wrote that AI "has to be part of the story" because several people are using AI tools to find bugs and build reproducible test cases. Vondra says he has nothing against her and likes her blog posts. But he says the claim does not match his understanding of why the reverts happened.

One data point in that debate is not disputed. PostgreSQL's August 2026 patch for version 18 carried 28 CVEs, against a historical average of a couple per release, according to Christensen's post as cited by Vondra. PostgreSQL has no official AI contribution policy yet, though one is expected soon.

OpenClaw's audit: 23 confirmed, none critical

The most detailed primary document is OpenClaw's own write-up of a security audit run with Trail of Bits through OpenAI's Patch the Planet initiative, published on 5 October. Trail of Bits submitted 27 private repository advisories and 3 standalone hardening pull requests. Of the advisories, 24 carried severity ratings: 0 Critical, 2 High, 16 Medium and 6 Low. The other 3 were defense-in-depth findings that crossed no documented trust boundary and were not rated.

OpenClaw says every actionable issue has been repaired and all three hardening PRs were merged. The mechanics are worth noting because they describe the same pattern that shows up across agent tooling. Trail of Bits used Codex-assisted workflows to search for issues and draft fixes, then checked findings manually before sending them over. OpenClaw then reviewed each report against its trust model and release history.

The recurring bug class was permissions lost between steps. A request would arrive with limited access, then spawn follow-on work that no longer carried those limits. In other cases a check saw one name for a thing while the system later used another, affecting both user identities and feature names. Archive checks inspected only part of an archive before extraction, and file paths could change after approval. OpenClaw's stated fix is to bind approval to the exact file, identity or action that will be used. It also re-checks if anything changes afterward.

Across OpenClaw, follow-on work must not gain access simply because it lost the context of the original request.

That line, from the company's blog, is the closest thing to a design principle in the document. It is also a fair description of what AI-assisted review tends to surface. Not exotic memory corruption, but gaps between what a system checks and what it later does.

The disclosure problem

InfoQ reported on 5 October on an article by Anil Madhavapeddy, professor of computer science at Cambridge and a core maintainer of the OCaml compiler. He argues that AI agents can turn public clues about vulnerabilities into working exploits, and so reduce the value of embargoes. Madhavapeddy describes fixing a path-traversal bug and noticing probes in his live webserver logs "with the exact bug pattern just minutes after opening the PR to fix the issue."

InfoQ cites a supporting number. In a recent study, a GPT-4 agent exploited 87% of vulnerabilities in a 15-vulnerability benchmark when given CVE descriptions, against 7% without them. Adrian Mouat of Chainguard told InfoQ that just opening a PR puts users in a bad place. He said projects may be pushed to publish releases before the associated source code, which he says breaks the fundamentals of open source.

Nick Craig-Wood, creator and maintainer of rclone, gave InfoQ a concrete measure of the load. "In the first 10 years of the rclone project we received about 20 security disclosures through GitHub. We had to deal with over 40 in the last month!" The volume is the point. Whether every one of those reports is a real bug is a separate question the quote does not answer.

Madhavapeddy's suggested mitigations are private vulnerability discussions, faster continuous releases and rapid protocol-level mitigations. None of them is a patch for the underlying asymmetry between the person who finds a bug and the agent that reads about it.

Tools that review and tools that script

The kernel has its own agentic reviewer. Sashiko is a self-contained system for reviewing Linux kernel code changes. It uses kernel-specific prompts and a multi-stage verification protocol, and works from a local git checkout or from mailing lists and git forges. Its README points maintainers at a configuration guide for mailing list tracking, maintainer notes and email policies. It notes that automated review replies go out from a separate address. The project says it is exploring coverage of other open-source projects and asks interested parties to contact Roman Gushchin.

Sashiko's local mode reads each commit in a range straight from git, without a scratch clone or patch application, and leaves the working tree untouched. The privacy note is blunt: patch data and surrounding files and history go to whichever LLM provider you configure. A --resume flag replays recorded requests, so an interrupted run only sends what was never answered.

On the other side of the same trend, Lunatik 5.0 lets scripts written in Lua run inside the Linux kernel. They reach character devices, Netfilter, XDP and TC hooks, kprobes, fsnotify, sched_ext, eBPF maps, sockets, netlink, crypto and kernel threads through Lua modules. It requires Linux 6.6 or later. The project's own example is a character device driver that generates random printable ASCII characters. It is loaded with sudo lunatik run passwd and removed with sudo lunatik stop passwd. A REPL ships with it, and its lines execute in the kernel.

The security implications are not spelled out on the project page, and this piece will not invent them. What can be said is that the barrier to writing kernel-resident code keeps dropping, in parallel with the barrier to finding bugs in it.

What the numbers do and do not show

A kernel datapath paper accepted at the eBPF'26 workshop in Prague, published on arXiv, illustrates how much performance work now runs through eBPF. Daniel Borkmann and Paul Chaignon describe netkit, an eBPF-based datapath that redirects packets between network namespaces to skip redundant backlog queue traversals. With minimal changes to the Cilium network plugin for Kubernetes, they report throughput improvements of up to 37% and parity between container-to-container and process-to-process communication.

The paper appeared in the proceedings of the 4th Workshop on eBPF and Kernel Extensions, pages 83 to 89, on 29 September. It is a performance result, not a security one, and should be read that way.

Meanwhile the patch treadmill keeps turning. Kagi said it will stop developing the Orion browser for Linux and Windows. The Linux beta no longer receives updates from the company after 2 October 2026, and Kagi warned users not to rely on it as a main browser. That is a maintenance and security-update story as much as a browser story. It follows the company's stated reason: a small team funded entirely by users, and no Chromium base to lean on.

Microsoft, by contrast, has made WSL Containers generally available through the latest WSL update. The release brings the wslc.exe command-line tool and a Windows API for launching Linux containers. According to How-To Geek, the feature was shown at Build 2026 and entered public preview in June. Compose support is the top request and is not there yet. A built-in container runtime changes the attack surface discussion on Windows desktops. But neither Microsoft's documentation nor the coverage quantifies that, so treat it as an open question.

Why the CVE count is the number to watch

The kernel's own disclosure history offers the clearest evidence that volume, not novelty, is the near-term problem. The Register reported on 5 October that Debian's latest kernel security update addresses 1,313 vulnerabilities. GIGAZINE covered the same figure while noting limits to tracking individual flaws one by one. That is the practical version of the argument Vondra is making about Postgres and that Madhavapeddy is making about embargoes: the bottleneck is human attention applied to a growing queue.

Against that backdrop, the 9.8 TCP flaw patched by Red Hat and Ubuntu on 3 November is a reminder of what still gets through. A use-after-free in IPv4 TCP, an in-the-wild scanning report on 2 November, and fixes landing the next day. Intersect had already urged Cardano operators to patch it on 31 October, before the distributor fixes were out.

None of these threads proves that AI is the cause of the kernel's security load. The OpenClaw audit is a vendor's account of its own engagement, and the Postgres reverts are contested by a maintainer with direct involvement. What the record does show is a set of projects experimenting with automated review under human sign-off. It also shows a maintainer culture that pushes back hard when a causal story gets ahead of the evidence, and a distribution pipeline that still has to ship the patches either way.

Comments 0

Sources

10
  1. 01AI Agents Are Disrupting Open Source Security DisclosureEN
  2. 02OpenClaw Completes Security Audit Through OpenAI's Patch the Planet InitiativeEN
  3. 03Postgres: Are we reverting patches because of bugs found by AI?EN
  4. 04Sashiko: Agentic review of Linux kernel codeEN
  5. 05Lunatik 5.0: scripting the Linux kernel with Lua, now with eBPF integrationEN
  6. 06Netkit: Specializing Linux Packet Delivery for Container NetworksEN
  7. 07Kagi Ends Orion Browser Development for Linux, Will Open-Source ItEN
  8. 08Linux containers now run natively on Windows via WSLEN
  9. 09Windows Developers Get a Built-In Linux Container Runtime: WSL ContainersEN
  10. 10Altera Quartus Linux jtagd bug fixesEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.