{
  "api_version": "1",
  "updated": "2026-11-10",
  "latest_event": "2026-11-09",
  "rights": "© Frontier Log, all rights reserved.",
  "story": {
    "slug": "openai-linux-kernel-exploits",
    "title": "OpenAI's Models Find More Linux Exploits for Less",
    "description": "OpenAI's models went from 24 Linux root exploits in June to 948 in one October run. It cost more than a July run, but each exploit cost under half as much.",
    "url": "https://frontierlog.ai/stories/openai-linux-kernel-exploits",
    "date": "2026-11-07",
    "image": "https://frontierlog.ai/og/story-openai-linux-kernel-exploits.png",
    "event": "2026-11-04-openai-patches-948-linux-kernel-bugs",
    "tags": [
      "cybersecurity"
    ],
    "steps": [],
    "tables": [
      {
        "title": "Linux local privilege escalation exploits, as reported by OpenAI",
        "unit": "",
        "rows": [
          {
            "event": "2026-06-22-daybreak",
            "date": "2026-06-22",
            "provider": "openai",
            "value": 24,
            "printed": null,
            "source": "https://openai.com/index/patch-the-planet",
            "where": "section 'Operating systems', 'Linux Kernel': \"GPT-5.5-Cyber identified security-relevant components across more than 30 million lines of code, flagged potential security issues, and then validated them dynamically, generating 8 kernel pointer information leak proof-of-concepts (PoCs) and 24 local privilege escalation exploits.\""
          },
          {
            "event": "2026-07-09-gpt-5-6",
            "date": "2026-07-09",
            "provider": "openai",
            "value": 166,
            "printed": null,
            "source": "https://openai.com/index/a-thousand-fewer-paths-to-root",
            "where": "section 'Stronger models find more vulnerabilities': \"in late July 2026, GPT-5.6 Sol found and exploited 166 unique LPE vulnerabilities for an end-to-end $513,690 cost in API credits.\""
          },
          {
            "event": "2026-11-05-gpt-6-cyber",
            "date": "2026-11-05",
            "provider": "openai",
            "value": 948,
            "printed": null,
            "source": "https://openai.com/index/a-thousand-fewer-paths-to-root",
            "where": "section 'Stronger models find more vulnerabilities': \"In the second scan, in mid-October 2026, the newer cyber model found and exploited 948 unique LPE vulnerabilities for an end-to-end $1,213,430 cost in API credits.\" The model is named GPT-6 Cyber in https://openai.com/index/introducing-gpt-6-cyber: \"Our researchers used GPT-6 Cyber to develop 948 privilege escalation exploits in Linux\"."
          }
        ]
      },
      {
        "title": "API cost per run, as reported by OpenAI",
        "unit": "$",
        "rows": [
          {
            "event": "2026-07-09-gpt-5-6",
            "date": "2026-07-09",
            "provider": "openai",
            "value": 513690,
            "printed": null,
            "source": "https://openai.com/index/a-thousand-fewer-paths-to-root",
            "where": "section 'Stronger models find more vulnerabilities': \"in late July 2026, GPT-5.6 Sol found and exploited 166 unique LPE vulnerabilities for an end-to-end $513,690 cost in API credits.\""
          },
          {
            "event": "2026-11-05-gpt-6-cyber",
            "date": "2026-11-05",
            "provider": "openai",
            "value": 1213430,
            "printed": null,
            "source": "https://openai.com/index/a-thousand-fewer-paths-to-root",
            "where": "'Findings' summary: \"The total run cost $1,213,430 in API credits, or $1,279 per LPE exploit.\""
          }
        ]
      }
    ],
    "cites": [
      {
        "url": "https://openai.com/index/patch-the-planet",
        "by": "OpenAI",
        "medium": "blog post",
        "date": "2026-06-22"
      },
      {
        "url": "https://openai.com/index/a-thousand-fewer-paths-to-root",
        "by": "OpenAI",
        "medium": "blog post",
        "date": "2026-11-04"
      },
      {
        "url": "https://openai.com/index/introducing-gpt-6-cyber",
        "by": "OpenAI",
        "medium": "blog post",
        "date": "2026-11-05"
      },
      {
        "url": "https://cognition.com/blog/swe-2-flash",
        "by": "Cognition",
        "medium": "blog post",
        "date": "2026-11-04"
      },
      {
        "url": "https://www.kimi.ai/blog/kimi-k3-1-pro",
        "by": "Moonshot AI",
        "medium": "blog post",
        "date": "2026-10-24"
      }
    ],
    "sources": [
      "https://openai.com/index/patch-the-planet",
      "https://openai.com/index/a-thousand-fewer-paths-to-root",
      "https://openai.com/index/introducing-gpt-6-cyber",
      "https://cognition.com/blog/swe-2-flash",
      "https://www.kimi.ai/blog/kimi-k3-1-pro"
    ],
    "updates": [],
    "body": [
      {
        "figure": "main"
      },
      {
        "p": "OpenAI said on Nov 4 that one run of its vulnerability research pipeline produced [948 working Linux local privilege escalation exploits](https://openai.com/index/a-thousand-fewer-paths-to-root), each one a way for an unprivileged user to get root on a stock desktop Linux system. The run took place in mid-October with what the post called \"an upcoming cyber model\". The next day, OpenAI's [launch post for GPT-6 Cyber](https://openai.com/index/introducing-gpt-6-cyber) named the model: \"Our researchers used GPT‑6 Cyber to develop 948 privilege escalation exploits in Linux\"."
      },
      {
        "table": {
          "cols": [
            "Model",
            "When",
            "Exploits",
            "Run cost",
            "Cost per exploit"
          ],
          "rows": [
            [
              "[GPT-5.5 Cyber](/model/gpt-5-5-cyber)",
              "[June](https://openai.com/index/patch-the-planet)",
              "24",
              "Not stated",
              "Not stated"
            ],
            [
              "[GPT-5.6 Sol](/model/gpt-5-6-sol)",
              "[July](https://openai.com/index/a-thousand-fewer-paths-to-root)",
              "166",
              "$513,690",
              "$3,094"
            ],
            [
              "[GPT-6 Cyber](/model/gpt-6-cyber)",
              "[October](https://openai.com/index/a-thousand-fewer-paths-to-root)",
              "948",
              "$1,213,430",
              "$1,279"
            ]
          ]
        }
      },
      {
        "h": "From 24 exploits to 948 in four months"
      },
      {
        "p": "OpenAI reported Linux kernel exploits on Jun 22, when it launched [Patch the Planet](https://openai.com/index/patch-the-planet), a Daybreak program that helps open-source maintainers fix vulnerabilities. GPT-5.5 Cyber had identified security-relevant components across more than 30 million lines of code and generated 24 local privilege escalation exploits. OpenAI noted that the model identified hundreds of issues, and that the 24 were the subset for which it generated exploits automatically."
      },
      {
        "p": "OpenAI then built the work into a pipeline that finds, reproduces and exploits kernel bugs and generates candidate patches for them. Human researchers review the results and report the vulnerabilities to maintainers alongside the patches. Its [Nov 4 post](https://openai.com/index/a-thousand-fewer-paths-to-root) gives two runs of that pipeline on the latest kernel. In late July, GPT-5.6 Sol found and exploited 166 unique vulnerabilities. In mid-October, GPT-6 Cyber found and exploited 948, which OpenAI's GPT-6 Cyber post calls \"more than five times the 166 exploited by GPT‑5.6 Sol using the same research harness\"."
      },
      {
        "h": "Each exploit cost less than half as much"
      },
      {
        "p": "OpenAI reported the API cost of both pipeline runs. The July run with GPT-5.6 Sol cost $513,690. The October run with GPT-6 Cyber cost $1,213,430, which OpenAI put at $1,279 per exploit. The October run cost 2.4 times as much as the July run and produced 5.7 times as many exploits. Divided the same way, the July run cost $3,094 per exploit, so the cost of each exploit fell by 59%. Both figures are the cost of the whole run, from scanning the kernel to developing the exploits. OpenAI gave no cost for the June work."
      },
      {
        "p": "OpenAI [called its costs](https://openai.com/index/a-thousand-fewer-paths-to-root) \"more an upper bound\", since it focused on measuring the model's capability rather than its cost, and expects that \"sufficient harness engineering could meaningfully reduce the cost at each stage\"."
      },
      {
        "h": "What the counts measure"
      },
      {
        "p": "The July and October counts come from the same harness on the latest kernel at the time, and OpenAI gives both as unique vulnerabilities. For the October run, it matched variants of the same bug: before that matching, the run produced 1,169 working exploits. The June count came from the earlier work, before the pipeline was built, so it was not measured the same way. OpenAI's GPT-6 Cyber post says it has disclosed 304 critical findings to the Linux maintainers with candidate patches, and is working to get the other vulnerabilities fixed."
      },
      {
        "h": "Two other labs limited what they released"
      },
      {
        "p": "Two other labs said in the same weeks that they had limited what they released because of cyber risk. Cognition said on Nov 4 that training raised [SWE-2 Flash's](https://cognition.com/blog/swe-2-flash) ExploitBench score from 11.3% to 34.8%, and that \"we are not sharing the weights publicly yet\" because the same skills can be used by attackers. It is building a program to share the weights with trusted industry partners and research labs."
      },
      {
        "p": "Moonshot AI said on Oct 24 that it removed cyber-related data from mid-training and the RL environments of [Kimi K3.1 Pro](https://www.kimi.ai/blog/kimi-k3-1-pro), its open-weight model. On its in-house benchmark, xdev-bench-hard, a checkpoint trained without those cuts solved 90.5% of tasks at pass@3, and the released model 81.0%. Moonshot said K3.1 Pro still \"improves substantially over its predecessor\" on cyber tasks, and that it aims \"not to release open weights that would extend the absolute frontier of cyber capability\"."
      }
    ],
    "events": [
      "2026-10-24-kimi-k3-1-pro",
      "2026-11-04-openai-patches-948-linux-kernel-bugs",
      "2026-11-04-swe-2-flash",
      "2026-11-05-gpt-6-cyber"
    ],
    "stories": [
      "ai-lab-bug-bounties"
    ],
    "models": [
      "gpt-5-5-cyber",
      "gpt-5-6-sol",
      "gpt-6-cyber",
      "swe-2-flash",
      "kimi-k3-1-pro"
    ]
  }
}