Trail of Bits uses Codex to backport fixes for 149 kernel CVEs

Trail of Bits published a writeup of an agentic workflow — built on OpenAI's Codex with GPT-5.6 — that backports security fixes from upstream into older stable releases. As of October 29 it had delivered fixes for 149 distinct CVEs into stable Linux kernel branches, plus "several" Fedora and Ubuntu package fixes. The relevant beat here isn't the model; it's the maintenance gap it targets.

The problem

Stable kernels and distro packages routinely never receive fixes that land only on the latest tree, because backporting is grinding manual work. Trail of Bits quotes XFS developer Dave Chinner calling it "effectively a full time job." The post's own census of the backlog, as of September 18, 2026:

Linux kernel CVEs unfixed in stable branches
  5.10.y   2,006
  5.15.y   1,889
  6.1.y    1,329
  6.6.y    1,012
  6.12.y     686
  6.18.y     337

Debian 13     1,128 source packages / 4,562 CVEs / 1,822 (39.9%) high or critical
Ubuntu 26.04  1,028 source packages /   770 CVEs /   335 (43.5%) high or critical

How the workflow runs

Agents identify fixes missing from a stable branch, adapt the upstream patch, assemble validation evidence, run tests, and prepare a submission packet. Engineers review every patch and submit it to maintainers — the model doesn't push to anyone's tree. Trail of Bits puts human effort at roughly 30 minutes for a clean backport and about three hours for an involved one, excluding per-project setup. Results are tracked on a public dashboard, and the effort runs under its Patch the Planet program; no tool or agent framework was released as a repo.

Caveats worth keeping

The post is unusually candid about its own numbers, and the caveats are the reason a practitioner should read it before citing the figures:

• Time estimates cover patch preparation and review only — "regression testing and maintainer acceptance remain essential."
• The CVE tables "are not an exact bug count"; they both over- and under-count, and are not a measure of exploitability.
• Distro figures are metadata findings, not individually source-confirmed vulnerabilities.
• The first kernel batch was fixes that applied cleanly; harder backports are still in progress, and the post gives no overall success rate.

Why it matters

Stable-branch rot is a real supply-chain exposure — the fix exists upstream, but the LTS kernel or distro package your fleet actually runs never gets it. If AI-assisted backporting holds up under maintainer review at scale, it attacks that gap directly. The skepticism to hold onto: this is a security vendor describing its own program, the acceptance rate is the number that matters and isn't published yet, and an LLM adapting a security patch is exactly the place a subtle, silent regression hides. The dashboard and the Bugzilla/mailing-list trail are the parts you can verify today; the "149 CVEs" headline is an output count, not a claim that 149 stable branches are now safe.