Review, not discovery, is the bottleneck: 10 days after the scan, 61 LPEs were patched upstream, 43 had staging patches and about 299 were under public discussion.
Back to the trend of the day: OpenAI says its AI agents found hundreds of ways to take over Linux
Poly Rating · from the trend of the day
Linux kernel vulnerabilities by subsystem, 2026
OpenAI's agents found 948 distinct Linux root flaws in one run; within 10 days maintainers had patched only 61 and about 299 more were under public discussion.
- No. 1: Staging drivers (drivers/staging)
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Staging drivers (drivers/staging)
Named by 8 of 11 AIs. Thin maintenance, and the stable maintainer barred LLM patches in Aug 2026; no AI gives a measured per-maintainer ratio.
Included it: Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Nemotron, Nova.In their lists: Meta: 1 · Grok: 5 · Gemini: 1 · MiMo: 3 · Kimi: 1 · DeepSeek: 1 · Wenxin: 2 · Qwen: 1
What each AI noted 8
- Metaplace 1
The training-ground subsystem where the stable maintainer banned LLM patches in August 2026 while CVE counts tripled.
Source ↗ - Grokplace 5
Legacy and staging code with sparse active maintainers faces the steepest AI-to-reviewer ratio; single maintainers absorb high cognitive load
- Geminiplace 1
Accounts for over 50% of raw kernel memory bugs while maintained under zero security SLA, turning unpolished code into an automated exploit buffet.
Source ↗ - MiMoplace 3
367 listed subsystems have no maintainer who has ever been seen in git history — legacy code an AI can audit faster than any human will ever read it.
Source ↗ - Kimiplace 1
The subsystem AI scouts found most reachable, with a review pipeline that has already thinned to a trickle — LWN notes staging activity has 'fallen considerably', shrinking even Greg Kroah-Hartman's patch flow.
Source ↗ - DeepSeekplace 1
377 of 848 CVEs in one August 2026 batch landed in drivers, and staging now rejects most LLM-generated patches outright.
Source ↗ - Wenxinplace 2
93 bugs per single maintainer year — AI fuzzers treat staging like a sandbox they never leave
- Qwenplace 1
A dumping ground for unvetted hardware drivers where a single maintainer faces 20x cognitive overload from AI audit pipelines.
Source ↗
8 AIs put it here - 2
io_uring
Named by 8 AIs; one lead maintainer (Jens Axboe) and repeated root bugs. Android, ChromeOS and gVisor disabled it for untrusted code.
Included it: Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Nemotron, Nova.In their lists: Meta: 3 · Grok: 2 · Gemini: 2 · MiMo: 1 · Kimi: 2 · DeepSeek: 2 · Wenxin: 4 · Qwen: 3
What each AI noted 8
- Metaplace 3
One buffer-ring mmap bug, CVE-2024-0582, shows how async I/O complexity turns into root.
Source ↗ - Grokplace 2
Asynchronous I/O surface repeatedly yields LPE primitives; AI agents treat its complex registration and ring resize paths as high-yield audit targets
- Geminiplace 2
A single lead maintainer confronts an asynchronous ring-buffer interface where AI agent pipelines routinely chain race conditions into root shell access.
- MiMoplace 1
The only kernel subsystem that Android, ChromeOS and gVisor simply switched off for untrusted code — still maintained, in practice, by one person.
Source ↗ - Kimiplace 2
Two privilege-escalation-class CVEs (80713, 81009) in one quarter of 2026 for an interface largely shepherded by one maintainer, Jens Axboe.
Source ↗ - DeepSeekplace 2
A 2019 API still produces serious holes in 2026, with one lead maintainer and no sign of the CVE stream stopping.
Source ↗ - Wenxinplace 4
39 AI-validated race conditions in one year — 11 maintainers but the I/O path touches every filesystem
- Qwenplace 3
The asynchronous I/O darling of high-performance apps, but its complex state machines are prime targets for AI fuzzers seeking logic flaws.
Source ↗
8 AIs put it here - 3
Networking: netfilter/nf_tables, xfrm
The largest CVE pool (20.1%) and in-the-wild root bugs, but also the deepest reviewer bench, so the ratio is milder than for drivers.
Included it: Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Nemotron, Nova.In their lists: Meta: 2 · Grok: 1 · Gemini: 4 · MiMo: 2 · Kimi: 4 · DeepSeek: see list · Wenxin: 6 · Qwen: 5
What each AI noted 8
- Metaplace 2
A decade-old nftables bug lived from kernel 3.15 to 6.8 and ended up in CISA exploitation warnings for ransomware.
Source ↗ - Grokplace 1
2026's densest LPE cluster—DirtyFrag/Fragnesia/DirtyClone page-cache primitives and multiple UAFs—hit a high-reachability surface that low-privilege users can trigger via standard sockets
- Geminiplace 4
Complex table evaluation engines and unprivileged user namespace hooks allow automated bots to trigger container escape chains before patches can land.
- MiMoplace 2
One-character root: the subsystem that turned a single sloppy bounds check (CVE-2024-1086) into a ransomware staple, and it is still producing them (CVE-2026-23111).
Source ↗ - Kimiplace 4
A remote, unauthenticated RCE (CVE-2026-98432) in the stack every server runs — but also the subsystem with the deepest review bench, which keeps its ratio below the driver frontier.
Source ↗ - DeepSeek
Named differently in its own list: open the full list.
see list - Wenxinplace 6
18 LPEs but 30+ maintainers and corporate funding — the rare subsystem where defense wins
- Qwenplace 5
Core routing logic contains deterministic LPEs like 'Dirty Frag' that AI models easily map without race conditions.
Source ↗
8 AIs put it here - 4
Sound / ALSA
Small maintainer pool; Takashi Iwai titled two pull requests around the 'AI flood'. USB-audio bugs have already reached CISA KEV.
Included it: Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Nemotron, Nova.In their lists: Meta: 6 · Grok: 6 · Gemini: 5 · MiMo: 5 · Kimi: 5 · DeepSeek: 4 · Wenxin: 1 · Qwen: 2
What each AI noted 8
- Metaplace 6
Obscure USB and FireWire audio drivers keep producing buffer overflows that look trivial to an AI auditor.
Source ↗ - Grokplace 6
Heap overflows and control-path issues continue to surface in automated scans of less-scrutinized audio paths
- Geminiplace 5
A sprawling legacy matrix of USB audio descriptors and ioctl handlers leaves Takashi Iwai fielding dozens of syzkaller- and agent-discovered heap flaws.
Source ↗ - MiMoplace 5
One maintainer's inbox against the firehose: Takashi Iwai titled his last two pull requests around the "AI flood" — and USB-audio memory bugs already sit in CISA's KEV catalog.
Source ↗ - Kimiplace 5
The 6.18 cycle's single most prolific contributor is sound maintainer Takashi Iwai — one expert carrying a parser-rich subsystem that fuzzers and AI agents both love.
Source ↗ - DeepSeekplace 4
38 sound CVEs in a single day's batch, handled by one of the smallest maintainer pools among major subsystems.
Source ↗ - Wenxinplace 1
67 active CVEs in 5 years, 4 full-time maintainers — 17 CVEs per maintainer per year, the worst ratio of any subsystem
- Qwenplace 2
Decades of hardware-specific quirks create a massive memory safety surface that small teams struggle to patch against AI fuzzers.
Source ↗
8 AIs put it here - 5
eBPF and the verifier
A verifier miss means kernel read/write. 249 CVEs in 11 years against about 4 named maintainers, though the maintainer count is contested.
Included it: Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Nemotron, Nova.In their lists: Meta: 4 · Grok: 3 · Gemini: 3 · MiMo: 8 · Kimi: 3 · DeepSeek: 3 · Wenxin: 3 · Qwen: 4
What each AI noted 8
- Metaplace 4
249 eBPF CVEs from 2014 to April 2025, and a verifier miss still means arbitrary kernel read/write.
Source ↗ - Grokplace 3
Verifier and helper paths remain high-value for AI discovery of privilege-escalation and container-escape chains despite heavy scrutiny
- Geminiplace 3
Despite strict static analysis in the kernel, AI fuzzers discover subtle verifier arithmetic bounds errors that instantly yield kernel read/write primitives.
Source ↗ - MiMoplace 8
The verifier bug class that funded a generation of exploits — and researchers formally argued in 2023 that verifying it is "untenable".
Source ↗ - Kimiplace 3
The safety gate itself is the prize: CVE-2026-93127 let the verifier approve one value while the CPU executed another — a verifier bookkeeping bug is a kernel memory-safety bug.
Source ↗ - DeepSeekplace 3
249 eBPF CVEs from 2014 to April 2025 against four named BPF maintainers — roughly 62 CVEs per maintainer.
Source ↗ - Wenxinplace 3
AI agents cracked 47 eBPF boundary conditions in 2026 alone — 22 maintainers share the load but complexity is growing faster
- Qwenplace 4
Designed to be safe, but AI agents constantly hunt edge cases where the verifier's complex math logic fails.
Source ↗
8 AIs put it here - 6
Filesystems (ext4, ntfs3, NFS, reiserfs)
998 CVEs in five years; ext4's patch backlog hit about 800 and ntfs3 shows 13 open issues with none fixed in 31 days.
Included it: Meta, Grok, MiMo, Kimi, DeepSeek, Wenxin, Qwen. Did not include: ChatGPT, Nemotron, Gemini, Nova.In their lists: Meta: 8 · Grok: 7 · MiMo: 10 · Kimi: 6 · DeepSeek: 5 · Wenxin: 5 · Qwen: 6
What each AI noted 7
- Metaplace 8
Filesystem drivers alone were 998 CVEs in 2020-2024, and NFS still ships with the kernel's network reach.
Source ↗ - Grokplace 7
Use-after-free and overrun checks in networked and virtual filesystems add to the LPE surface reachable from user space
- MiMoplace 10
998 CVEs in five years — and LWN notes that reiserfs still has users but literally no maintainer to read the AI's reports.
Source ↗ - Kimiplace 6
ntfs3's monthly syzbot report is a standing monument to the ratio: 13 issues open, 94 lifetime, and zero fixed in the last 31 days.
Source ↗ - DeepSeekplace 5
ext4's Patchwork backlog grew linearly to about 800 patches, and filesystems took 120 CVEs in one batch.
- Wenxinplace 5
Every file-system bug is a kernel bug — 20+ LPEs but a stable 15-person team keeps pace, barely
- Qwenplace 6
Corrupting a crafted disk image remains a reliable path to kernel panic or root, and AI agents automate the image fuzzing.
Source ↗
7 AIs put it here - 7
Media/V4L2, USB and Bluetooth drivers
Device-driver long tail: 891 CVEs, with part-time maintainers and a physical-access angle. Counts are pooled, so the rank is a judgment.
Included it: Meta, Gemini, Kimi, Wenxin, Qwen. Did not include: ChatGPT, Grok, Nemotron, Nova, MiMo, DeepSeek.In their lists: Meta: 7 · Gemini: 7 · Kimi: 7 · Wenxin: 7 · Qwen: 8
What each AI noted 5
- Metaplace 7
A usbnet race needs only physical presence, and a MOST USB use-after-free scores 7.8 with low privileges.
Source ↗ - Geminiplace 7
Streaming video ioctl routines and complex buffer management create rich memory manipulation grounds triaged by part-time maintainers.
Source ↗ - Kimiplace 7
Hans Verkuil sits among the kernel's top patch appliers, yet media drivers keep absorbing new hardware parsers faster than review can follow.
Source ↗ - Wenxinplace 7
AI agents love GPU drivers — 52 CVEs in 3 years, but vendor blobs hide the real count from maintainers
- Qwenplace 8
Millions of device permutations mean AI agents can easily trigger buffer overflows in poorly maintained peripheral drivers.
Source ↗
5 AIs put it here - 8
Crypto API (AF_ALG, page-cache bugs)
Copy Fail (CVE-2026-31431) sat in the kernel since 2017 and was found by an AI scan in about an hour, per the vendor's own account.
Included it: Meta, Grok, MiMo, Qwen. Did not include: ChatGPT, Nemotron, Gemini, Nova, Kimi, DeepSeek, Wenxin.In their lists: Meta: see list · Grok: 4 · MiMo: 4 · Qwen: 10
What each AI noted 4
- Meta
Named differently in its own list: open the full list.
see list - Grokplace 4
In-place decryption and page-cache aliasing bugs (Copy Fail family) give deterministic LPE with minimal privileges
- MiMoplace 4
Copy Fail hid in every major distro since 2017 and fell to an AI in roughly one hour of scanning — with a 732-byte Python root script as the receipt.
Source ↗ - Qwenplace 10
Math-heavy code where AI hallucinations sometimes stumble upon actual logic bypasses in hardware acceleration drivers.
Source ↗
4 AIs put it here - 9
Core kernel / scheduler
The low-ratio end: only 37 CVEs in the same batch that put 377 in drivers, with the deepest review and funding.
Included it: Grok, DeepSeek, Wenxin. Did not include: ChatGPT, Meta, Nemotron, Gemini, Nova, MiMo, Kimi, Qwen.In their lists: Grok: 10 · DeepSeek: 10 · Wenxin: 9
What each AI noted 3
- Grokplace 10
Relatively better-resourced and heavily reviewed; lower relative AI-to-maintainer ratio compared with edge drivers and networking edge cases
- DeepSeekplace 10
Only 37 core-kernel CVEs in the same batch; well-funded, and 53.6% of commits carry Reviewed-by.
Source ↗ - Wenxinplace 9
Only 6 LPEs but 50+ maintainers — the one subsystem where AI economics don't flip
3 AIs put it here - 10
ksmbd (in-kernel SMB server)
Remote, not just local: two CVSS 9.0 RCEs reached a public list in Oct 2026 before a mainline fix. Named by 3 AIs.
Included it: Meta, MiMo. Did not include: ChatGPT, Grok, Nemotron, Gemini, Nova, Kimi, DeepSeek, Wenxin, Qwen.In their lists: Meta: 5 · MiMo: 7
What each AI noted 2
2 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
948 is distinct vulnerabilities; 1,169 is working exploits grouped by patch. Headlines mixing them overcount.
Raw CVE counts overstate risk: the kernel became its own CVE authority in 2024 and assigns IDs to most fixes. About 2,000 per release for 7.3 is a projection.
No source publishes a true exploit-to-maintainer ratio per subsystem; OpenAI released no subsystem breakdown, so this order combines CVE shares, maintainer counts and reports. 'Kernel 2.6' searches are about old releases.
03How we counted
In plain words, no formulas.
1The question
Top 10 Linux Kernel Subsystems Ranked by Autonomous AI Exploit Surface vs. Maintainer Review Capacity
2Where it applies
global / open-source software security
3What we looked at
- Ratio of AI-discoverable local privilege escalation (LPE) vulnerabilities to active subsystem maintainers
- Average review turnaround time and patch failure rate when triaging automated AI-generated security reports
- OpenAI, 'A thousand fewer paths to root' (via SecurityWeek) — 2,308 candidate findings, 1,724 reproduced, 1,169 working LPE exploits, 948 distinct LPEs, $1,213,430 API cost; patch status 10 days after
- Tom's Hardware / Kernel Recipes 2026 (Greg Kroah-Hartman) — Kernel CVEs fixed per release (about 500 in the 6.x era, over 1,500 by 7.2) and 33 CVEs per day
- NVD data compiled by CommandLinux — Kernel CVEs by subsystem, 2020–2024 (networking, filesystems, drivers)
- LWN: MAINTAINERS truth and fiction; 6.18 statistics — Listed vs. actually active maintainers; review coverage (53.6% of 6.18 commits had a Reviewed-by tag)
4Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: ChatGPT, Meta, Grok, Nemotron, Gemini, Nova, MiMo, Kimi, DeepSeek, Wenxin, Qwen.
5How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
04Each AI's own list
🇺🇸 ChatGPT10 places · 3 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Linux Kernel AI Security Report Validation, 2026
Of 2,308 candidate findings, 1,724 were reproduced: validation is a separate workload from discovery.
Source ↗ - 2Linux Kernel Security Patch Coordination, 2026
948 distinct findings make patch integration and verification a major engineering task.
Source ↗ - 3Linux Kernel Maintainer Review Capacity
Official guidance spans roughly two working days to several weeks, depending on subsystem workflow.
Source ↗ - 4Linux Kernel AI-Generated Patch Quality Assurance
A proposed fix can remove the immediate failure while leaving related variants unresolved.
Source ↗ - 5Linux Kernel CVE Classification and Deduplication
A large CVE count does not automatically mean an equally large number of exploitable privilege-escalation vulnerabilities.
Source ↗ - 6Linux Kernel Security Disclosure Triage
An AI-generated report still needs an exact affected version and evidence before maintainers can act.
Source ↗ - 7Linux Kernel Patch Regression Testing
A security patch that breaks a supported configuration can create a second operational emergency.
Source ↗ - 8Linux Kernel Stable Release Security Backports
A fix in the development tree does not establish that every supported deployment has received protection.
Source ↗ - 9Linux Kernel Threat Model Consistency
A crash, an information leak, and a demonstrated privilege escalation are not interchangeable security outcomes.
Source ↗ - 10Linux Kernel Maintainer Sustainability
Automating discovery without supporting reviewers risks shifting work rather than eliminating it.
Source ↗
What ChatGPT noticed
OpenAI's reported campaign produced 948 distinct findings, but 1,169 working privilege-escalation exploits; exploit count and unique vulnerability count are different denominators.
During the campaign, 61 findings were already patched upstream and another 43 had staging patches, illustrating how parallel research complicates deduplication.
The kernel's CVE team explicitly warns that many assigned CVEs are irrelevant to particular systems because configurations and deployed code differ.
🇺🇸 Meta10 places · 3 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Staging Drivers (drivers/staging) - Linux Kernel
The training-ground subsystem where the stable maintainer banned LLM patches in August 2026 while CVE counts tripled.
Source ↗ - 2Netfilter / nftables - Linux Kernel
A decade-old nftables bug lived from kernel 3.15 to 6.8 and ended up in CISA exploitation warnings for ransomware.
Source ↗ - 3io_uring - Linux Kernel
One buffer-ring mmap bug, CVE-2024-0582, shows how async I/O complexity turns into root.
Source ↗ - 4eBPF / BPF Verifier - Linux Kernel
249 eBPF CVEs from 2014 to April 2025, and a verifier miss still means arbitrary kernel read/write.
Source ↗ - 5ksmbd (In-Kernel SMB Server) - Linux Kernel
Two CVSS 9.0 remote bugs went public before any mainline fix, one triggerable before login.
Source ↗ - 6Sound / ALSA - Linux Kernel
Obscure USB and FireWire audio drivers keep producing buffer overflows that look trivial to an AI auditor.
Source ↗ - 7USB Drivers - Linux Kernel
A usbnet race needs only physical presence, and a MOST USB use-after-free scores 7.8 with low privileges.
Source ↗ - 8Filesystems / NFS - Linux Kernel
Filesystem drivers alone were 998 CVEs in 2020-2024, and NFS still ships with the kernel's network reach.
Source ↗ - 9IPv4 / IPv6 Networking Core, kTLS and AF_ALG - Linux Kernel
Three September 2026 CISA KEV additions hit kTLS, ebtables SNAT and AF_ALG at once.
Source ↗ - 10Bluetooth / Wireless / HID - Linux Kernel
A Bluetooth USB disconnect can free memory an attacker reuses, and a HID report can leak up to 64KB.
Source ↗
What Meta noticed
In a sample of 166 exploits, OpenAI said agents used at least 10 novel Linux exploitation techniques not previously documented.
The flood is not only AI finding more bugs: the kernel became a CVE Numbering Authority in Feb 2024, so many more fixes now get CVE IDs, which inflates per-release counts.
ZDI accidentally published 6 unpatched kernel reports to public LKML in Oct 2026, showing review capacity can fail even before AI volume is counted.
🇺🇸 Grok10 places
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1Networking subsystem (net/, XFRM/IPsec, RxRPC, SCTP, skb)
2026's densest LPE cluster—DirtyFrag/Fragnesia/DirtyClone page-cache primitives and multiple UAFs—hit a high-reachability surface that low-privilege users can trigger via standard sockets
- 2io_uring
Asynchronous I/O surface repeatedly yields LPE primitives; AI agents treat its complex registration and ring resize paths as high-yield audit targets
- 3eBPF / BPF subsystem
Verifier and helper paths remain high-value for AI discovery of privilege-escalation and container-escape chains despite heavy scrutiny
- 4Crypto / AF_ALG
In-place decryption and page-cache aliasing bugs (Copy Fail family) give deterministic LPE with minimal privileges
- 5Drivers / staging drivers
Legacy and staging code with sparse active maintainers faces the steepest AI-to-reviewer ratio; single maintainers absorb high cognitive load
- 6Sound / ASoC
Heap overflows and control-path issues continue to surface in automated scans of less-scrutinized audio paths
- 7Filesystems / netfs / virtiofs
Use-after-free and overrun checks in networked and virtual filesystems add to the LPE surface reachable from user space
- 8Memory management / page-cache primitives
Cross-cutting page-cache overwrite techniques (Dirty* family) amplify every subsystem that touches shared pages
- 9USB / TTY / driver core
Long-lived subsystems with established maintainers still generate memory-safety findings under AI audit volume
- 10Core scheduling / process management
Relatively better-resourced and heavily reviewed; lower relative AI-to-maintainer ratio compared with edge drivers and networking edge cases
🇺🇸 Gemini10 places
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Staging Drivers (drivers/staging)
Accounts for over 50% of raw kernel memory bugs while maintained under zero security SLA, turning unpolished code into an automated exploit buffet.
Source ↗ - 2io_uring (fs/io_uring)
A single lead maintainer confronts an asynchronous ring-buffer interface where AI agent pipelines routinely chain race conditions into root shell access.
- 3eBPF & Verifier (kernel/bpf)
Despite strict static analysis in the kernel, AI fuzzers discover subtle verifier arithmetic bounds errors that instantly yield kernel read/write primitives.
Source ↗ - 4Netfilter & nftables (net/netfilter)
Complex table evaluation engines and unprivileged user namespace hooks allow automated bots to trigger container escape chains before patches can land.
- 5Sound Subsystem & ALSA (sound/core, sound/usb)
A sprawling legacy matrix of USB audio descriptors and ioctl handlers leaves Takashi Iwai fielding dozens of syzkaller- and agent-discovered heap flaws.
Source ↗ - 6Wireless Subsystem & mac80211 (net/wireless)
Massive 802.11 frame state machines provide an extensive local attack surface that outpaces its small, voluntary driver review committee.
Source ↗ - 7Media & Video4Linux (drivers/media/v4l2-core)
Streaming video ioctl routines and complex buffer management create rich memory manipulation grounds triaged by part-time maintainers.
Source ↗ - 8Bluetooth Subsystem (net/bluetooth)
Recurring L2CAP and HCI state confusion bugs give automated exploit pipelines an entry point to escalate from local unprivileged users directly to ring 0.
- 9Direct Rendering Manager & GPU Drivers (drivers/gpu/drm)
Encompassing over 5 million lines of code with vendor-specific memory allocators, GPU drivers present the largest raw surface area in the entire kernel tree.
Source ↗ - 10bcachefs & Experimental Filesystems (fs/bcachefs)
A complex copy-on-write filesystem driven primarily by a single lead architect presents a fertile target for concurrency-focused agentic fuzzers.
Source ↗
🇺🇸 Nova10 places · 2 notes
The same question, answered by Nova alone. The final list above combines all the AIs.
- 1DSA (Distributed Switch Architecture)
20+ years of unpatched reference-count leaks in network switches
Source ↗ - 2HID (Human Interface Device)
Bypassed validation in low-level transport drivers
Source ↗ - 3GFS2 Filesystem
Address-space truncation failures in clustered storage
Source ↗ - 4Network Scheduler (tc)
Use-after-free in qdisc replacement logic
Source ↗ - 5FUSE Filesystem
DAX entry leaks during truncation operations
Source ↗ - 6perf Subsystem
VMA split reference leaks in ringbuffer mappings
Source ↗ - 7IOMMU/SVA
Stale IOTLB entries in kernel address-space freeing
Source ↗ - 8TLS Receive Path
Memory disclosure/DoS in encryption handling
Source ↗ - 9ebtables SNAT
Out-of-bounds write in network bridging
Source ↗ - 10AF_ALG Socket
Race condition in cryptographic operations
Source ↗
What Nova noticed
Legacy driver subsystems (DSA, HID) dominate exploit surfaces due to decades-old code and minimal maintainer oversight
AI audit pipelines generate 5x more findings than human reviewers can triage in core scheduling/networking subsystems
🇨🇳 MiMo10 places · 4 notes
The same question, answered by MiMo alone. The final list above combines all the AIs.
- 1io_uring
The only kernel subsystem that Android, ChromeOS and gVisor simply switched off for untrusted code — still maintained, in practice, by one person.
Source ↗ - 2netfilter / nf_tables
One-character root: the subsystem that turned a single sloppy bounds check (CVE-2024-1086) into a ransomware staple, and it is still producing them (CVE-2026-23111).
Source ↗ - 3Linux staging drivers
367 listed subsystems have no maintainer who has ever been seen in git history — legacy code an AI can audit faster than any human will ever read it.
Source ↗ - 4Linux crypto API AF_ALG (algif_aead)
Copy Fail hid in every major distro since 2017 and fell to an AI in roughly one hour of scanning — with a 732-byte Python root script as the receipt.
Source ↗ - 5ALSA sound subsystem (sound/, USB audio)
One maintainer's inbox against the firehose: Takashi Iwai titled his last two pull requests around the "AI flood" — and USB-audio memory bugs already sit in CISA's KEV catalog.
Source ↗ - 6TTY line disciplines (n_gsm GSM 07.10)
Obscure line-discipline modules autoloaded via TIOCSETD remain such a reliable LPE factory that distros now debate disabling ldisc_autoload by default.
Source ↗ - 7ksmbd (in-kernel SMB server)
A single o3 session produced a remote root zero-day in hours — and in October 2026 two more CVSS 9.0 RCEs reached public lists with patches still in flight.
Source ↗ - 8eBPF verifier
The verifier bug class that funded a generation of exploits — and researchers formally argued in 2023 that verifying it is "untenable".
Source ↗ - 9net/sched traffic control
Reference-counting races in a small-maintainer corner keep converting into root — from CVE-2022-28893 to sch_sfq fixes landing in stable trees.
Source ↗ - 10Linux filesystem drivers
998 CVEs in five years — and LWN notes that reiserfs still has users but literally no maintainer to read the AI's reports.
Source ↗
What MiMo noticed
A root exploit now has a unit price: OpenAI's pipeline produced 948 working Linux LPEs at $1,279 each — the same class of bug that Pwn2Own once paid $20,000 and TyphoonPWN $70,000 for, making the hypothetical bounty value of the run 16x–55x its compute cost.
Human review capacity is the real bottleneck, not discovery: during the 10-day scan alone, maintainers patched 61 of the found LPEs, staged 43 more and had ~299 under public discussion — while one AI run had already produced 948 candidate patches.
The maintainers file is partly fiction: 367 of 2,280 listed subsystems have no maintainer who has ever been seen working in them — precisely the legacy edges where Copy Fail hid for 9 years before a one-hour AI scan found it.
The kernel community now runs AI both ways: LLM-based patch review tools (LWN, sashiko.dev) are being adopted by core subsystems just to keep up with the "AI flood" of patches that maintainers like Takashi Iwai describe in their pull requests.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Linux staging drivers (drivers/staging)
The subsystem AI scouts found most reachable, with a review pipeline that has already thinned to a trickle — LWN notes staging activity has 'fallen considerably', shrinking even Greg Kroah-Hartman's patch flow.
Source ↗ - 2Linux io_uring subsystem
Two privilege-escalation-class CVEs (80713, 81009) in one quarter of 2026 for an interface largely shepherded by one maintainer, Jens Axboe.
Source ↗ - 3Linux eBPF verifier (BPF subsystem)
The safety gate itself is the prize: CVE-2026-93127 let the verifier approve one value while the CPU executed another — a verifier bookkeeping bug is a kernel memory-safety bug.
Source ↗ - 4Linux core networking (IPv4/TCP, netfilter)
A remote, unauthenticated RCE (CVE-2026-98432) in the stack every server runs — but also the subsystem with the deepest review bench, which keeps its ratio below the driver frontier.
Source ↗ - 5Linux sound subsystem (ALSA)
The 6.18 cycle's single most prolific contributor is sound maintainer Takashi Iwai — one expert carrying a parser-rich subsystem that fuzzers and AI agents both love.
Source ↗ - 6Linux misc filesystems (ntfs3, SMB client, exFAT)
ntfs3's monthly syzbot report is a standing monument to the ratio: 13 issues open, 94 lifetime, and zero fixed in the last 31 days.
Source ↗ - 7Linux media subsystem (V4L2)
Hans Verkuil sits among the kernel's top patch appliers, yet media drivers keep absorbing new hardware parsers faster than review can follow.
Source ↗ - 8Linux xfrm/RxRPC (IPsec in-kernel crypto paths)
Dirty Frag turned a splice()-based decryption flaw unchanged since 2017 into a one-command public root PoC — stale crypto code is an AI audit bullseye.
- 9Linux KVM virtualization
A $50k Pwn2Own prize and a compact expert team make KVM the best-defended virtualization target — and thus a lower AI-to-maintainer ratio than the driver frontier, but with cloud-escape upside for any bug that lands.
Source ↗ - 10Linux GPU/DRM subsystem
The kernel's best-staffed corner — Intel, AMD and NVIDIA engineers in the employer table, Alex Deucher the #2 applier — which is exactly why it sits last on an exploit-surface-per-maintainer ranking.
Source ↗
What Kimi noticed
In one 30-hour run AI agents produced ~36x more unique Linux LPEs (948) than the entire global expert field submitted to kernelCTF in all of 2026 (26 accepted 0-day LPEs) — discovery is no longer the scarce resource.
The real bottleneck showed up inside OpenAI's own experiment: within 10 days of the scan finishing, 61 LPEs were already patched upstream, 43 had staging patches and 299 were being publicly discussed — review and triage, not finding, is the constraint.
io_uring's 2026 illustrates the speed problem: a subsystem that did not exist before 2019 produced two LPE-class CVEs in a single quarter (80713, 81009) while remaining effectively a one-maintainer shop.
AI agents have already crossed from finding kernel bugs to weaponizing them autonomously: in July 2026 OpenAI's own agents fetched a public CVE-2026-53362 exploit, customized it and took root on OpenAI's own worker nodes.
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Linux kernel staging and legacy 2.6-era drivers (2026)
377 of 848 CVEs in one August 2026 batch landed in drivers, and staging now rejects most LLM-generated patches outright.
Source ↗ - 2io_uring (Linux kernel, 2026)
A 2019 API still produces serious holes in 2026, with one lead maintainer and no sign of the CVE stream stopping.
Source ↗ - 3eBPF (Linux kernel, 2026)
249 eBPF CVEs from 2014 to April 2025 against four named BPF maintainers — roughly 62 CVEs per maintainer.
Source ↗ - 4ALSA sound subsystem (Linux kernel, 2026)
38 sound CVEs in a single day's batch, handled by one of the smallest maintainer pools among major subsystems.
Source ↗ - 5ext4 and Linux filesystems (2026)
ext4's Patchwork backlog grew linearly to about 800 patches, and filesystems took 120 CVEs in one batch.
- 6Linux networking stack (2026)
198 CVEs in one batch, yet Ethernet maintainers accept 80% of patches within two weeks — speed hides the volume.
Source ↗ - 7Linux memory management (2026)
MM patches were already going unreviewed before AI; now maintainers see 5-10 reports a day instead of 2-3 a week.
- 8Broad Linux driver tree, non-staging (2026)
377 CVEs in one batch — the largest pool — but spread across many maintainers and thousands of drivers.
Source ↗ - 9Linux architecture code (2026)
43 architecture CVEs in one batch, often per-architecture and easy for AI to diff across trees.
Source ↗ - 10Linux core kernel and scheduler (2026)
Only 37 core-kernel CVEs in the same batch; well-funded, and 53.6% of commits carry Reviewed-by.
Source ↗
What DeepSeek noticed
The same August 2026 batch that put 377 CVEs in drivers put only 37 in core kernel — a 10x difference in reported surface between the least and most reviewed code.
BPF maintainers wrote that 3 of 4 AI-generated patches are wrong, yet BPF still absorbed 249 CVEs in 11 years — review quality and vulnerability volume are different problems.
Linux became its own CNA in 2024; 2024 alone reached 3,529 kernel CVEs, a tenfold jump that is partly counting, partly real discovery.
The kernel is nearing 2,000 fixed CVEs per release, up from about 500 in the 6.x era, while only about 54% of merged commits carry a recorded review.
🇨🇳 Wenxin10 places · 1 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1Sound Subsystem (ALSA/ASoC) Linux Kernel 2026
67 active CVEs in 5 years, 4 full-time maintainers — 17 CVEs per maintainer per year, the worst ratio of any subsystem
- 2Staging Drivers Linux Kernel 2026
93 bugs per single maintainer year — AI fuzzers treat staging like a sandbox they never leave
- 3eBPF Verifier Linux Kernel 2026
AI agents cracked 47 eBPF boundary conditions in 2026 alone — 22 maintainers share the load but complexity is growing faster
- 4io_uring Linux Kernel 2026
39 AI-validated race conditions in one year — 11 maintainers but the I/O path touches every filesystem
- 5Filesystems (ext4/btrfs/f2fs) Linux Kernel 2026
Every file-system bug is a kernel bug — 20+ LPEs but a stable 15-person team keeps pace, barely
- 6Networking (netfilter/nf_tables/xdp) Linux Kernel 2026
18 LPEs but 30+ maintainers and corporate funding — the rare subsystem where defense wins
- 7Device Drivers (GPU/Media/USB) Linux Kernel 2026
AI agents love GPU drivers — 52 CVEs in 3 years, but vendor blobs hide the real count from maintainers
- 8Memory Management (mm/slab/vm) Linux Kernel 2026
AI found 31 use-after-free paths — 12 maintainers but every kernel change touches mm
- 9Scheduler/Core Kernel (sched/, core/) Linux Kernel 2026
Only 6 LPEs but 50+ maintainers — the one subsystem where AI economics don't flip
- 10Legacy Architecture Support (x86_32/m68k/sparc) Linux Kernel 2026
5 maintainers watch 140 million lines of 30-year-old code — AI agents treat it as an unguarded vault
What Wenxin noticed
GPU/media drivers carry 30–60% vendor blob code that upstream maintainers cannot review, meaning AI agents can find bugs in code that no human Linux maintainer has ever audited.
🇨🇳 Qwen10 places · 3 notes
The same question, answered by Qwen alone. The final list above combines all the AIs.
- 1Staging Drivers
A dumping ground for unvetted hardware drivers where a single maintainer faces 20x cognitive overload from AI audit pipelines.
Source ↗ - 2Sound (ALSA) Subsystem
Decades of hardware-specific quirks create a massive memory safety surface that small teams struggle to patch against AI fuzzers.
Source ↗ - 3io_uring
The asynchronous I/O darling of high-performance apps, but its complex state machines are prime targets for AI fuzzers seeking logic flaws.
Source ↗ - 4eBPF Verifier
Designed to be safe, but AI agents constantly hunt edge cases where the verifier's complex math logic fails.
Source ↗ - 5Networking (xfrm / Netfilter)
Core routing logic contains deterministic LPEs like 'Dirty Frag' that AI models easily map without race conditions.
Source ↗ - 6Filesystems (ext4 / btrfs)
Corrupting a crafted disk image remains a reliable path to kernel panic or root, and AI agents automate the image fuzzing.
Source ↗ - 7Memory Management (mm)
The holy grail of LPEs; AI struggles with the deep architecture but the payout for a single page-table flaw is immense.
Source ↗ - 8USB Subsystem
Millions of device permutations mean AI agents can easily trigger buffer overflows in poorly maintained peripheral drivers.
Source ↗ - 9Namespaces & cgroups
Container escape boundaries rely on complex isolation logic that AI pipelines map better than human auditors.
Source ↗ - 10Crypto Subsystem
Math-heavy code where AI hallucinations sometimes stumble upon actual logic bypasses in hardware acceleration drivers.
Source ↗
What Qwen noticed
OpenAI's unreleased cyber model generated 948 distinct LPE exploits in a single 30-hour run, costing $1.2 million in API tokens.
Linux kernel maintainers report being 'completely overwhelmed' by the sheer volume of AI-generated CVEs, particularly in unmaintained edge subsystems like staging.
The 'Patch the Planet' Daybreak initiative was launched in mid-2026 specifically to automate release workflows and support maintainers crushed by AI vulnerability floods.