OpenAI Found 948 Kernel Bugs. Only 77 US Profiles Say Kernel Engineer
AI now finds Linux kernel bugs faster than humans can review the fixes. Here is how to find kernel reviewers through Reviewed-by tags, MAINTAINERS, and lore.
On Nov 4, OpenAI said one automated run had produced patches for 948 Linux privilege-escalation bugs, and every one of those patches still needs a human to review it. If you hire kernel talent for a cloud or Android company, the person you need is the one who can say "this fix is right" in a way maintainers trust. That person rarely has "kernel" in their job title, and you won't find them on a GitHub contribution graph.
Why kernel review, not kernel bugs, is now the bottleneck
AI can now find kernel bugs and draft fixes faster than the community can review them, so trusted reviewers are the scarce skill. OpenAI's run went from 2,308 candidate findings to 1,724 reproduced vulnerabilities to 948 validated patches, at a total API bill of $1,213,430. That works out to about $1,280 per vulnerability.
What happened next shows where things slow down. Within 10 days, reports counted 61 bugs patched upstream, 43 with staged patches, and 299 under public discussion. Those 100-plus fixes were made by other contributors working in parallel during the same window, so it's wrong to credit all 104 to OpenAI. Either way, only about 11% of the findings had been fixed or staged after 10 days. The rest had not landed.
Review also catches mistakes. In an early run, 20 of 200 exploits got root through a different flaw from the one they were assigned, and review caught it. CIQ, which publicly tracks how many of the OpenAI patches get accepted, says only a fraction of AI-generated kernel patches can ship without changes. CIQ's point is that once review takes longer than the patch saved, the work has simply moved from writing code to reviewing it.
The kernel's own numbers point the same way:
| Release | Commits with Reviewed-by | Note |
|---|---|---|
| 6.19 | 56% | Start of the decline |
| 7.0 | 54% | |
| 7.2 | 48% | 1,111 Assisted-by commits, up from 301 in 7.1 |
| 7.3 | 50.9% (9,545 of 18,765) | Busiest release ever |
A Reviewed-by tag is a line at the bottom of a commit message saying a named person read the patch and vouches for it. An Assisted-by tag marks a commit written with help from an AI tool. In a single cycle, AI-assisted commits nearly quadrupled while review coverage fell below half.
That ratio explains why one new reviewer is worth more than one new author. Every patch that lands needs a reviewer. If you hire authors and no reviewers, your patches pile up unreviewed.
What Lorenzo Stoakes stepping down actually signals
Stoakes stepping back shows that reviewer capacity in core subsystems like memory management is shrinking just as demand rises. Around Nov 10 (some sources say Nov 11), Lorenzo Stoakes of Arm said he would stop reviewing the mm core, reclaim, and THP (transparent huge pages) areas.
His reason, in his own words: "Given the current throughput in the kernel resulting from AI-generated series, I unfortunately lack the bandwidth to generally be able to offer review except in those areas I maintain."
One correction to the usual framing: his decision was not a response to OpenAI's campaign, and you shouldn't present it as one. The two events are separate symptoms of the same pressure. Neither caused the other.
What matters for hiring is what he kept. Stoakes has contributed to the kernel since 2014 and wrote The Linux Memory Manager. He kept his maintainer roles and gave up only his reviewer roles. In the MAINTAINERS file (the kernel's directory of who owns each subsystem), M: lines list maintainers and R: lines list designated reviewers. The R: lists for mm core, reclaim, and THP now have one fewer name.
A reviewer leaving an R: line is a better hiring signal than any job title change.
For companies that depend on mm, such as cloud providers and Android vendors, this is a real gap. Arm's role for Stoakes is the model to copy: a company funds reviewer time directly, and the subsystem keeps a reviewer it would otherwise lose.
Why a title search misses more than 99% of kernel-skilled profiles
Searching by job title finds fewer than 1 in 100 of the people who list kernel skills, and some of the matches work on the wrong kernel. In Refolk's index of professional profiles, 13,208 US profiles list "Linux Kernel" as a skill. Only 77 have a kernel-engineer or kernel-developer title.
Here's why. Kernel engineers usually carry generic titles like Staff Engineer or Software Engineer. Their real credentials are trailer lines in git log and entries in MAINTAINERS, and a title filter can't see either.
The skill tag has the opposite problem. In a 25-profile sample of the US skill-tagged group, nearly every top title was an executive: CTOs, co-founders, chief architects. The tag mostly tells you someone once worked on the kernel, not that they review patches today.
The title search also returns wrong matches. Among the 77 US kernel-titled profiles, Apple was the most common employer (9 in a 25-person sample), followed by Google (3) and Oracle (3). Apple's kernel is XNU, not Linux, so many of those Apple "Kernel Engineers" are probably not Linux reviewers. That's an inference you should check profile by profile, but it means a title filter both misses the right people and returns the wrong ones.
| Metric | Value | Source |
|---|---|---|
| US profiles with "Linux Kernel" skill | 13,208 | Refolk's index |
| US profiles with kernel engineer/developer title | 77 | Refolk's index |
| Title share of skill group | 0.58% (1 in 172) | Derived |
| Germany profiles with "Linux Kernel" skill | 1,110 | Refolk's index |
| US vs Germany ratio | 11.9x | Derived |
Germany is the second market worth knowing. Its 1,110 skill-tagged profiles are about a twelfth of the US pool, and Intel was the top named employer in the sample. If you're hiring in Europe, Intel is the first employer to map.
Where kernel reviewers actually leave evidence
Kernel reviewers leave a public, timestamped record in commit trailers, the MAINTAINERS file, and mailing-list threads on lore.kernel.org. This is how to find kernel developers who review, as opposed to people who once listed the skill.
The signals, from strongest to weakest:
- Reviewed-by trailers in mainline commits. Each one means the person read the patch and the maintainer accepted their judgment.
- R: and M: entries in MAINTAINERS. These are official, subsystem-specific, and visible to anyone with a clone of the tree.
- lore.kernel.org review threads. lore.kernel.org is the public archive of kernel mailing lists. Reviews that never made it into a merged commit, including pushback that killed a bad patch, only show up there.
- Per-release statistics from LWN. LWN publishes the top reviewers by tag count for each release.
Look at who tops those lists and you'll see something important. In 7.1, Konrad Dybcio added his tag to 334 changes and Dmitry Baryshkov to 300, almost all on Qualcomm drivers and devicetree files written by Qualcomm colleagues. In the 70-day 6.19 cycle, Charles Keepax reviewed 310 commits, more than four a day, mostly sound drivers.
So reviewers usually already work at a vendor and review that vendor's code. Commit authorship won't show you who they are, because they aren't the authors. To find the reviewer for a vendor's subsystem, look at who replies to that vendor's patches on lore.
The hard part is connecting a lore email address to a person you can actually contact. A mailing-list address doesn't tell you where someone works today, and their profile shows a generic title at a company you might not guess. Bridging that gap is what Refolk does: you describe the person in plain English, with the subsystem, the trailer, and the vendor, and get back a ranked shortlist of matched profiles.
A five-step workflow to source kernel memory management engineers
To source kernel memory management engineers, start with the code and work out to the person. Don't start with the profile.
- Map the subsystem. Read the MAINTAINERS entries for mm core, reclaim, and THP. Write down every M: and R: name. This is your top tier.
- Mine the trailers. Run
git log --since="2 years ago" -- mm/and pull out every Reviewed-by. Count them by name. The people with dozens of tags who aren't in MAINTAINERS are your best targets: trusted, but not yet tied to an official role. - Read lore for the gaps. Search lore.kernel.org for mm threads and note who gives substantive review on AI-assisted series. The R: lists for these areas just lost a name, so whoever is covering that work right now is the person to watch.
- Note the employer pattern. Reviewers cluster by vendor. Someone who reviews every patch from one company almost certainly works there.
- Resolve identities. Match the email domain, name, and subsystem to a current profile. Expect the title to say "Software Engineer." Don't filter that out.
Step five is where most searches stall. A linux kernel recruiter has to confirm that a lore address belongs to the same person as a profile titled "Staff Engineer," find out whether they've changed employers since their last trailer, and check whether they'd consider moving. Repeat that for every name on an mm shortlist and the search drags on, which is why so many of these roles end up handed to an agency.
How to pitch a kernel reviewer differently from an author
Pitch reviewers on protected review time and upstream standing, not on shipping features. A reviewer's reputation is the trust maintainers place in their judgment, and they won't want a role that spends that trust on rushing your patches through.
What works:
- Funded upstream review time. Arm's role for Stoakes is the template. Put a fixed share of the week in writing.
- A named subsystem. "Review mm and reclaim for our fleet" beats "join the kernel team."
- A filter in front of them. OpenAI's Patch the Planet program, run with Trail of Bits and HackerOne, pays for review before reports reach maintainers. Offer your reviewer the same thing: triage support so they only see patches worth their attention.
- Honest scope on AI. Given Stoakes's post, any reviewer you approach is already dealing with a flood of AI-generated series. Tell them how your team uses AI and who checks the output before it reaches them.
What fails: a message that cites their GitHub stars, or an offer to "own the roadmap" that never mentions upstream work. GitHub activity is a weak signal for this group. If you cite a specific Reviewed-by on a specific patch, you show you did the research their trailers make possible.
FAQ
How do I find kernel developers who aren't on GitHub?
Use the kernel's own records. Pull Reviewed-by trailers from git log on the subsystem you care about, read the R: and M: lines in MAINTAINERS, and search lore.kernel.org for who reviews patches in that area. Then match those names and email addresses to current profiles. Kernel review happens on mailing lists, so GitHub activity is a weak signal for this group.
Is a "Linux Kernel" skill tag a useful filter?
It's a starting point, not a filter. In Refolk's index, 13,208 US profiles list the skill, but a sample of them was dominated by CTOs and co-founders who did kernel work earlier in their careers. Only 77 have a kernel title, and the most common employer among those is Apple, whose kernel is XNU. Check every candidate against commit trailers before you reach out.
How many of OpenAI's 948 kernel patches have been merged?
Within 10 days of the Nov 4 announcement, reports counted 61 bugs patched upstream, 43 with staged patches, and 299 under public discussion. Those 104 fixed or staged bugs were worked on by other contributors in parallel too, so not all of them should be credited to OpenAI. CIQ is publicly tracking how many of the OpenAI patches are accepted, and says only a fraction of AI-generated kernel patches can ship without modification.
Did OpenAI's 948 bugs cause Lorenzo Stoakes to step down as a reviewer?
No. His post cited the general flood of AI-generated patch series, and he wasn't responding to OpenAI's campaign. The two events happened about a week apart and reflect the same pressure on human review, but you shouldn't present one as the cause of the other.
The data behind this
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.