OpenAI pipeline produces 948 working Linux root exploits

Summary

• An automated OpenAI pipeline generated 948 working local privilege-escalation exploits against the Linux kernel.
• The October 2026 run cost $1.21 million in API credits, roughly $1,279 per working exploit.
• OpenAI says an unprivileged account on an unpatched system now effectively equals root access.

The latest

An automated pipeline built by OpenAI's cyber team produced 948 working local privilege-escalation exploits against the Linux kernel, each letting an unprivileged user reach root on a widely used distribution. The company published the research on 4 November 2026, saying the system also found, verified and generated patches for the flaws, and ran on a cyber model that has not been released.

Details

• The run: In a mid-October 2026 run, the pipeline flagged 2,308 candidate findings and reproduced 1,724 vulnerabilities inside an isolated environment, according to OpenAI. Beyond the 948 privilege-escalation exploits, it produced 291 proofs of concept for information-disclosure flaws.
• The cost: OpenAI put the total bill at $1,213,430 in API credits, which works out to $1,279 per working exploit. The company framed the figure as the economics of the capability rather than a one-off research expense.
• Generational jump: According to OpenAI, the newer model found more than five times as many novel vulnerabilities as GPT-5.6 Sol did in a July 2026 run. That earlier run exploited 166 flaws at a cost of $513,690, a fraction of the October output.
• New techniques: Across a sample of 166 exploits, the agents used at least 10 previously unseen exploitation techniques, OpenAI said. The company withheld details of those techniques, citing safety concerns.
• Human review: Human researchers checked the pipeline's output before anything was reported to kernel developers, according to the company. OpenAI said it sent hundreds of reports and patches covering the critical flaws, and is releasing the remainder gradually with partners in the Patch the Planet initiative.
• Patch status: When scanning and analysis wrapped after 10 days, 61 vulnerabilities had been patched in the mainline kernel, 43 patches were in preparation, and 299 were in reporting or public discussion, OpenAI said. The company did not set a date for disclosing the rest.
• The warning: OpenAI assessed that, in the near and medium term, holding a low-privilege account on a system missing the latest patches should be treated as practically equivalent to root access against a well-resourced attacker. That framing shifts the risk calculation for any multi-user Linux deployment.
• The advice: The company recommended isolation and sandboxing layers such as Firecracker and gVisor to contain the blast radius of a successful exploit. It also urged organizations to rescan their codebases with each new model generation, rather than treating a clean audit as durable.

Background

Local privilege escalation turns a foothold into full control: an attacker who already has an ordinary account on a Linux machine uses a kernel flaw to become root, gaining the ability to read any data, alter the system and persist.

Between the lines

The $1,279-per-exploit figure is the part defenders have to price in. Capability that once required scarce expert labor now scales with a budget line, and the five-fold jump over July 2026 suggests the curve is steep. OpenAI's own recommendation to rescan code with every model generation concedes that a codebase audited today is not audited for long.

What's next

Watch how fast the 43 patches in preparation and 299 flaws still under public discussion reach the mainline kernel, and whether OpenAI's partners disclose the remaining reports on a published schedule.

Source: OpenAI