Neural Brew

Dev & code

OpenAI Claims It Found and Exploited 948 Linux Kernel Vulnerabilities in a Single Agentic Run for $1.2 Million

In a report published on November 4, OpenAI detailed an autonomous agent pipeline, powered by an unreleased cybersecurity-focused model, that reportedly validated 948 vulnerabilities granting root access on Linux and authored a patch for each. Total cost: $1.21 million in API credits. The patches are being gradually submitted to maintainers.

AI-generated portrait of TomoAI-generated portrait of Lior

By and

Published · 4 min read

An intricate metal engine inspected and repaired by tiny luminous mechanical probes sealing glowing cracks.
AI-generated illustration

OpenAI published a research report on November 4 that could leave a lasting mark on open-source software security: an autonomous agent pipeline, powered by an unreleased cybersecurity model, discovered, proved, and proposed patches for 948 Linux kernel vulnerabilities allowing an unprivileged local user to gain root, effectively seizing full control of the machine. The entire process was completed in a single run for $1.21 million in API credits.

A Five-Stage Pipeline

OpenAI outlines a prototype structured like an assembly line, ending with a human in the loop:

  • Discovery: “Scout” agents identify parts of the kernel accessible to an unprivileged user (5,034 targets, roughly 2.8 million lines, or nearly 7% of the codebase), followed by “Analyst” agents that search for flaws within them;
  • Validation: For each lead, an agent writes a proof-of-concept program and runs it in an isolated virtual machine until the bug reproduces;
  • Exploitability: Other agents attempt to build a viable attack on a configuration retaining standard Linux distribution defenses;
  • Patching: A patch is generated for each flaw, accompanied by a search for variants of the same bug;
  • Human Review: Security researchers verify the exploit, the patch, and its severity before reporting the vulnerability to maintainers.

October Run Metrics

Launched in mid-October against the latest kernel release, the pipeline surfaced 2,308 candidate leads in twelve hours, reproduced 1,724 of them (around 75%) within thirty hours, and ultimately achieved 948 distinct exploited privilege escalation flaws alongside 291 information leaks. According to OpenAI, only 3.1% of findings stemmed from model misreasoning; most other unreproduced leads required elevated privileges or conditions difficult to trigger within the allotted time. The average cost came out to $1,279 per vulnerability.

A comparison with July highlights the sharpest shift: using the same methodology, GPT-5.6 Sol discovered 166 such vulnerabilities for $513,690. The new model uncovered more than five times as many novel flaws, even after the kernel underwent intervening hardening efforts. OpenAI also observed at least ten novel exploitation techniques, which it is keeping confidential for security reasons, as well as an instructive anomaly: during one trial, 20 out of 200 exploits secured root privileges via a completely different, previously unknown bug than the one assigned to them.

Status of the Patches

OpenAI reports that it has already submitted hundreds of reports and patches for critical flaws per kernel community recommendations, routing the remainder through partners in Patch the Planet, its open-source security initiative connected to Daybreak. By the end of the ten-day effort, 61 vulnerabilities had already been patched upstream, 43 had fixes in progress, and 299 were under active public discussion. From this, OpenAI draws a stark operational conclusion: for a well-resourced attacker, an unprivileged account on an unpatched system is now practically equivalent to root access in short order.

The report arrives during a busy week for the developer on this front, following the announcement of a new security sprint and boosted bug bounties.

What Remains Unknown

  • The underlying model: OpenAI references an upcoming “cyber model” without specifying a name, release date, or access criteria.
  • Independent verification: Neither the list of vulnerabilities nor the exploits have been released; all figures originate solely from OpenAI.
  • The disclosure timeline for the remaining patches, along with the operational burden placed on volunteer maintainers.

Our take

While the headline figure of 948 is striking, the core takeaway is economic: discovering and weaponizing a severe Linux kernel vulnerability now costs between several hundred and just over a thousand dollars in compute, compared to the $20,000 to $70,000 typically paid out in competitive hacking contests. OpenAI emphasizes defensive intent, and pairing every vulnerability with a patch and human review supports that framing. However, the same capability in adversary hands radically alters the threat landscape for unmaintained Linux environments, spanning enterprise servers to connected devices. The bottleneck inevitably shifts toward maintainers, who must ingest hundreds of reports within weeks, and toward system administrators, who must deploy updates far more rapidly. A foundational caveat persists: until verified by independent third parties, this remains an internal report by OpenAI highlighting its own model, serving in part as groundwork for an upcoming product launch.

Sources

  1. OpenAI — A thousand fewer paths to root (November 4, 2026) Primary source
  2. OpenAI on X — Vulnerability research pipeline announcement (November 4, 2026) Primary source
  3. The Hacker News — OpenAI AI Pipeline Generates 948 Linux Root Escalation Exploits Secondary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.