OpenAI's vulnerability pipeline finds Linux root exploits at scale

Overview

Updated 53 minutes ago

On November 4, OpenAI described a pipeline that scanned the Linux kernel, reproduced 1,724 vulnerabilities, and produced 948 working local privilege-escalation exploits plus candidate patches. The run cost $1,213,430 in API credits, about $1,279 per exploit.

Linux runs on most of the world's servers and on Android phones. A local privilege-escalation flaw turns a low-privileged account into root, so finding hundreds in one pass changes how quickly defenders must patch. OpenAI says the pipeline also discovered at least 10 exploitation techniques it had not seen before.

Why it matters

If finding a root exploit costs $1,279, a low-privileged account on an unpatched Linux machine is effectively root.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

948

Working LPE exploits

Distinct local privilege-escalation exploits with candidate patches, from one mid-October run.

1,724

Vulnerabilities reproduced

Candidate findings that the pipeline validated dynamically in isolated virtual machines.

2,308

Candidate findings

Static scans of security-relevant kernel code produced these initial candidates.

$1,213,430

Total API cost

Cost of the full mid-October campaign in API credits.

$1,279

Cost per LPE exploit

The campaign cost divided by 948 distinct exploits.

~7%

Share of kernel code scanned

Scout agents selected 5,034 audit targets, about 7.8 million lines, as reachable by low-privileged users.

LPEs patched upstream during the scan

Linux maintainers patched these while the 10-day scan ran; 43 more had staging patches.

Novel exploitation techniques

Techniques OpenAI says the agents discovered and used, to the best of its knowledge.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Sam Altman

Chief executive of OpenAI

Leads OpenAI

Organizations Involved

OpenAI

Artificial intelligence company

Built and ran the vulnerability pipeline

OpenAI is an AI lab in San Francisco that published the Linux kernel findings on November 4, 2026.

Linux kernel maintainers

Open-source maintainer community

Reviewing reported bugs and candidate patches

The distributed group that reviews and merges changes to the Linux kernel.

kernelCTF

Vulnerability bounty program

Benchmark for OpenAI's adapted run

A Google-run program that pays researchers for working exploits against the Linux kernel.

Patch the Planet

OpenAI security initiative

Ongoing

OpenAI's effort to use its models to find and patch software vulnerabilities.

Timeline

June 2026 November 2026

5 events Latest: 5 days ago

Tap a bar to jump to that date

November 2026
OpenAI publishes 'A thousand fewer paths to root'
Latest Publication

The report details the pipeline, the 948 exploits, the $1.2 million cost and at least 10 novel exploitation techniques.
5 days ago
October 2026
New cyber model finds 948 Linux LPE exploits
Research

The pipeline scans 5,034 audit targets, reproduces 1,724 vulnerabilities, and generates 948 exploits and patches for $1,213,430.
October 2026
Linux maintainers patch 61 reported LPEs mid-scan
Patch

Another 43 have staging patches and 299 are being discussed publicly while OpenAI's scan runs.
October 2026
July 2026
GPT-5.6 Sol run finds 166 unique Linux LPEs
Research

The earlier campaign costs $513,690 in API credits and sets the baseline for the October run.
July 2026
June 2026
OpenAI's GPT-5.5 Cyber finds 24 Linux LPE exploits
Research

Patch the Planet's first public Linux results identify kernel vulnerabilities and build working exploits to validate them.
June 2026

Scenarios

Linux maintainers merge AI-generated patches at scale

Possible Resolves by Q2 2027

Discussed by: OpenAI's report and Phoronix coverage

OpenAI says human researchers reviewed the outputs and sent bugs and patches to maintainers. If subsystem maintainers accept the code, the 948 findings become fixes rather than disclosures. The trigger is patches entering the mainline kernel rather than staying on mailing lists.

Maintainers reject the AI patches as low quality

Possible Resolves by Q1 2027

Discussed by: Kernel mailing-list reviewers and Hacker News commenters

Kernel maintainers require reproducible steps, correct commit messages and no style regressions. If the candidate patches fail that review, the value of the run is the bug reports, not the code. A public rejection of a batch would set that precedent.

Rival labs replicate the pipeline and disclosure volume surges

Possible Resolves by End of 2027

Discussed by: Axios and SecurityWeek

OpenAI's cost figure gives other labs a target. If a second company fields a similar pipeline, kernel maintainers face a larger flow of machine-generated reports. OpenAI already observed 299 LPEs being discussed publicly during its own 10-day scan.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

August 2016

DARPA Cyber Grand Challenge (2016)

Seven teams brought autonomous systems to Las Vegas to find, exploit and patch software flaws without human help. The winning system, Mayhem, collected $2 million. The contest was designed to test whether machines could perform security work end to end.

Then

Mayhem won the $2 million prize and the contest demonstrated autonomous bug-finding in a controlled setting.

Now

The event became a reference point for automated vulnerability research, though commercial products took years to follow.

Why this matters now

OpenAI's pipeline is the same idea applied to a real operating system rather than a contest binary, with a published cost per exploit.

2015–present

syzkaller and syzbot (2015 onward)

Google engineer Dmitry Vyukov built syzkaller, a coverage-guided fuzzer for the Linux kernel. The syzbot service later ran it continuously and reported crashes to maintainers automatically. It surfaced thousands of kernel bugs over the following years.

Then

Kernel developers received a steady stream of automated crash reports, some of them serious.

Now

Continuous fuzzing became a standard part of kernel development, and maintainers adapted their triage process to machine-generated reports.

Why this matters now

OpenAI's run tests whether the same shift can happen for exploit development and patch writing, not just crash detection.

2007–present

Pwn2Own and kernel exploit bounties (2007 onward)

The Zero Day Initiative's Pwn2Own contest turned working exploits into priced commodities. OpenAI cites $20,000 from Pwn2Own and $70,000 from TyphoonPWN as historical bounties for Linux kernel local privilege-escalation exploits.

Then

Exploit researchers gained a public market for high-severity bugs.

Now

Bounty prices became a rough measure of what a working exploit is worth to defenders and attackers.

Why this matters now

OpenAI applies those bounty figures to its 948 exploits and estimates a hypothetical value of $18.9 million to $66.3 million against a $1.2 million cost.