{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "Blog · Opinion · 5 November 2026",
    "lede": "Two days ago I wrote that the models find bugs faster than people can fix them. OpenAI has now run a pipeline that finds, proves and writes the patch, 948 times, for about $1,279 each, and I think it moves the bottleneck rather than removing it.",
    "readingMinutes": 4,
    "words": 902
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/a-machine-that-writes-the-patch",
    "json": "https://www.anchorterminal.com/blog/a-machine-that-writes-the-patch.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/a-machine-that-writes-the-patch.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/a-machine-that-writes-the-patch.min.md"
  },
  "markdown": "By Vlad Cealicu, 5 November 2026. 902 words, about 4 min to read.\n\n- 948: unique local privilege escalations in the Linux kernel, each with a candidate patch\n- $1,279: per issue in API credits, $1.2 million for the run\n- 7%: of the kernel covered, about 2.8 million lines\n\nOne run, mid-October 2026, as OpenAI reported it. Sources [1] and [2].\n\nAt four in the morning on Thursday 5 November I read OpenAI's post about it, and I was still reading it an hour later. Two days ago I wrote [a post about the models giving us more work than we can check](/blog/a-machine-that-gives-us-work), and the example I used was security, where the models were finding vulnerabilities far faster than anyone could triage, report and patch them, and the bottleneck was people. This is a pipeline that does the patch as well.\n\n## What they ran\n\nIn mid-October OpenAI pointed an autonomous pipeline running an upcoming cyber model, one it hasn't named or released, at 5,034 targets in the Linux kernel, about 2.8 million lines or roughly 7% of it [[1]](#ref-1). It came back with 2,308 candidate findings within 12 hours, reproduced 1,724 of them within 30 hours, which is 74.7% of the candidates, and only 3.1% of the reproduced ones came from the model reading the code wrong [[1]](#ref-1) [[2]](#ref-2). It then built 1,169 working local privilege escalation exploits, the kind that turns an ordinary user on a machine into root, and wrote a distinct patch for 948 of them, which is the number OpenAI uses for the unique ones [[1]](#ref-1). The whole run cost $1,213,430 in API credits, about $1,279 per unique issue, and OpenAI says it found at least ten exploitation techniques it judged to be new, the details of which it has kept back [[1]](#ref-1) [[2]](#ref-2).\n\nFrom the kernel to a candidate patch\n\n| Stage | Count |\n| --- | --- |\n| Audit targets | 5034 |\n| Candidate findings | 2308 |\n| Reproduced | 1724 |\n| Working root exploits | 1169 |\n| Unique, each with a patch | 948 |\n\nEach series has a zero-based scale.\n\nThe price is the part I keep coming back to. A run in late July with GPT-5.6 Sol found 166 unique ones for $513,690, about $3,095 each, so in under three months the cost per issue more than halved [[1]](#ref-1). OpenAI also aimed the same pipeline at Google's kernelCTF setup and got 21 working exploits and 25 information leaks for $32,936 in API credits, about $1,568 an exploit, against a bounty programme that could pay something like $1.5 million if all 21 were accepted, which OpenAI is careful to say hasn't happened [[1]](#ref-1). OpenAI calls its costs upper bounds, because the run was measuring what the model could do, not trying to do it cheaply [[1]](#ref-1).\n\nAPI cost per unique root exploit, in dollars\n\n| Run | Cost |\n| --- | --- |\n| GPT-5.6 Sol, late July | 3095 |\n| The upcoming cyber model, mid-October | 1280 |\n\nEach series has a zero-based scale.\n\n## Where the bottleneck goes\n\nI think this is good news, and I'm pro AI on it, and the patch is the part that was missing. But a candidate patch isn't a merged patch, and the reporting is clear on it, \"the 948 figure is not a count of merged patches\" as Unite.AI put it [[1]](#ref-1). Of the earlier findings, 61 were patched upstream, 43 had patches in staging and 299 had been reported or discussed by the end of the scan, and researchers still review every exploit, patch and severity before anything goes out [[1]](#ref-1). OpenAI's own report says it plainly, \"Vulnerability discovery is becoming easier to scale, but prioritization and validation remain operational bottlenecks\" [[2]](#ref-2).\n\nThe people at the other end of this are already stretched. The kernel was fixing around 500 CVEs a release through most of the 6.x years, passed 1,000 with Linux 7.0 and 1,500 with 7.2, and is heading for 2,000, and Jakub Kicinski, who looks after kernel networking, wrote \"We are completely overwhelmed\" [[3]](#ref-3). In May Linus Torvalds said AI tools are great \"but only if they actually help, rather than cause unnecessary pain and pointless make-believe work\", and asked people to read the documentation, make a patch and add something beyond what the tool found [[4]](#ref-4). A finding that comes with a tested patch is what he asked for, so I think this is the better kind of AI report, and it's also 948 of them on top of everyone else's.\n\nLinux kernel CVEs fixed per release, at least\n\n| Release | CVEs |\n| --- | --- |\n| Linux 6.x, most releases | 500 |\n| Linux 7.0 | 1000 |\n| Linux 7.2 | 1500 |\n\nEach series has a zero-based scale.\n\nSo the checking I wrote about two days ago moves somewhere else. It used to be the patch that people were short of time for, and now it's reviewing a patch someone else wrote, at a rate nobody has planned for.\n\n## The part that needs care\n\nI'd be careful with one thing, and I don't think it takes anything away from the work. A pipeline that finds and proves a root exploit for $1,279 is as useful to an attacker as it is to a defender, and the only difference right now is who has the model. OpenAI hasn't released it, has kept the new techniques back, and is working through the Linux maintainers and its Patch the Planet partners to get fixes out first [[1]](#ref-1) [[2]](#ref-2). That's the right order, defenders first, and I think it only works for as long as the fixes land faster than the same capability turns up somewhere else, which brings it back to the people reviewing the patches.\n\nI don't know how long that window is. The cost halved in three months, the reports are already arriving faster than maintainers can read them, and the patches still need a person to say yes. If you maintain a part of the kernel and you're reading this, I'd like to know whether a report that comes with a proven exploit and a patch saves you time or costs you more of it.\n\n---\n\nMore in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md)\n\n- Older: [Unreliable compared to what?](https://www.anchorterminal.com/blog/unreliable-compared-to-what.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nRead on 5 November 2026. Quotes are as published. OpenAI's own report, \"A thousand fewer paths to root\", wouldn't load for me, so its figures here are as two outlets reported them on 4 and 5 November.\n\n[1] Unite.AI, 4 November 2026, \"OpenAI Pipeline Finds 948 Linux Kernel Privilege Escalation Flaws\", reporting OpenAI's \"A thousand fewer paths to root\". Source for the mid-October 2026 run on 5,034 targets (about 2.8 million lines, roughly 7% of the kernel), 2,308 candidates in 12 hours, 1,724 reproduced in 30 hours with 3.1% from incorrect model analysis, 1,169 working local privilege escalation exploits and 948 distinct patches, $1.213 million in API credits or $1,279 per unique issue, the late-July GPT-5.6 Sol run (166 for $513,690), the kernelCTF adaptation (21 exploits and 25 information leaks for about $32,936, and the roughly $1.5 million bounty being conditional), costs as upper bounds, 61 earlier findings patched upstream, 43 in staging and 299 reported or discussed, \"the 948 figure is not a count of merged patches\", human review of every exploit, patch and severity, and the ten new techniques being withheld. https://www.unite.ai/openai-pipeline-finds-948-linux-kernel-privilege-escalation-flaws/\n\n[2] SecurityOnline, 5 November 2026, \"OpenAI Cyber Model Uncovers 1,724 Linux Kernel Vulnerabilities and Builds 948 Root Exploits\". Source for the model being an upcoming OpenAI cyber model, 74.7% of candidates confirmed, the total of $1,213,430, at least ten new techniques, the Patch the Planet project, and the quote \"Vulnerability discovery is becoming easier to scale, but prioritization and validation remain operational bottlenecks\". The per-issue costs of $3,095 and $1,568 are my arithmetic from references 1 and 2. https://securityonline.info/openai-cyber-model-linux-kernel-vulnerabilities/\n\n[3] Tom's Hardware, 1 September 2026, \"Linux kernel nears 2,000 CVEs per release as AI bug hunters scour 40 million lines of code\". Source for roughly 500 CVEs per release through much of the 6.x era, more than 1,000 with Linux 7.0 and more than 1,500 with Linux 7.2, Greg Kroah-Hartman's figures from Kernel Recipes 2026, and Jakub Kicinski's \"We are completely overwhelmed\". https://tomshardware.com/software/linux/linux-kernel-nears-2-000-cves-per-release-as-ai-bug-hunters-scour-40-million-lines-of-code-maintainers-say-they-are-completely-overwhelmed\n\n[4] Security Boulevard, 18 May 2026, \"Torvalds Offers Guidance as AI Bug Reports Clog Up Linux Security Workflow\". Source for Linus Torvalds' \"AI tools are great, but only if they actually help, rather than cause unnecessary pain and pointless make-believe work\", and his guidance to read the documentation, create a patch and add value beyond what the AI found. https://securityboulevard.com/2026/05/torvalds-offers-guidance-as-ai-bug-reports-clog-up-linux-security-workflow/\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-05",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.7",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Opinion",
        "url": "https://www.anchorterminal.com/blog/opinion/"
      },
      {
        "name": "A machine that writes the patch too",
        "url": ""
      }
    ],
    "description": "Vlad Cealicu on OpenAI's autonomous pipeline that reproduced 1,724 Linux kernel vulnerabilities and wrote a candidate patch for each of 948 local privilege escalations, for about $1,279 each. Why it answers the bottleneck from the verification post, where the bottleneck moves next, and the part that needs care.",
    "facts": [
      "Vlad Cealicu",
      "5 November 2026"
    ],
    "h1": "A machine that writes the patch too",
    "image": "https://www.anchorterminal.com/assets/og/blog-a-machine-that-writes-the-patch.png",
    "path": "/blog/a-machine-that-writes-the-patch",
    "published": "2026-11-05",
    "section": "blog",
    "title": "OpenAI's 948 Linux root exploits, and who reviews the patches | Anchor Terminal",
    "toc": null,
    "updated": "2026-11-05",
    "url": "https://www.anchorterminal.com/blog/a-machine-that-writes-the-patch"
  },
  "tokens": {
    "markdown": 2400,
    "slim": 780
  },
  "version": 1
}