OpenAI said Wednesday that an agentic pipeline built on an upcoming cyber model reproduced 1,724 Linux kernel vulnerabilities and produced 948 working local privilege escalation exploits, at a cost of $1,279 per exploit. The company published the results as a research write-up rather than a vulnerability disclosure tied to a named flaw.
The Mid-October Run Processed 2,308 Candidate Kernel Findings
OpenAI described a five-stage pipeline. In a run in mid-October, it analyzed 2,308 candidate kernel findings and dynamically reproduced 1,724 of them in about 30 hours. It then wrote proof-of-concept code for 291 information leaks and produced 948 working local privilege escalation exploits. The target was the Linux kernel in a commonly used distribution.
The agents used at least 10 novel Linux exploitation techniques, according to the company. The total cost was $1,213,430 in API credits, which works out to $1,279 per privilege escalation exploit.
OpenAI Puts the Run at 30 Hours, Not the 10 Days Some Outlets Reported
Some outlets gave the length of the run as 10 days. OpenAI’s own figure is 30 hours, and that figure is used here.
The Pipeline Builds on 24 Exploits Found by GPT 5.5 Cyber in June
The work follows an earlier effort. OpenAI’s GPT 5.5 Cyber model found 24 local privilege escalation exploits in June 2026. The October run increased the output by a large factor, from 24 exploits to 948.
OpenAI Treats an Unprivileged Account as Equivalent to Root
OpenAI states that an unprivileged account on an unpatched operating system should be treated as equivalent to root access for a well-resourced attacker. The company argues that patch generation must keep pace with exploit generation, and it pairs the research with AI-generated patches. In OpenAI’s framing, automated patch writing is the counterweight to automated exploit writing, and the company presents the two as halves of the same effort, with neither able to run far ahead of the other.
The source did not name any specific CVE from the 1,724 reproduced vulnerabilities, so the write-up does not identify which kernel flaws were involved or whether they have been fixed.
The Cost Per Exploit Falls to Roughly $1,300
The headline figure is the price. At about $1,279 per working exploit, mass generation of kernel privilege escalation exploits is now priced in a range that a funded group could afford many times over.
UK NCSC and Google Data Frame the Disclosure Volume
The same reporting places the OpenAI results alongside two outside assessments. The UK National Cyber Security Centre published an assessment on Tuesday saying AI cyber tools have almost certainly increased zero-day discovery. Google Threat Intelligence Group figures show monthly disclosed vulnerabilities rising from 5,045 in January 2026 to 10,740 in August 2026.
Those numbers describe a rising volume of disclosures, and the OpenAI write-up adds a measure of how cheaply exploit development can follow discovery. The conversion rates inside the run are notable. Of 2,308 candidate findings, 1,724 were reproduced, roughly three in four, and 948 exploits were produced, meaning more than half of the reproduced bugs were turned into working privilege escalation code. The 291 information-leak proofs of concept show the pipeline also handled a separate class of bug.
The company’s argument that patching must keep pace rests on the gap between the speed of automated exploit creation and the speed at which maintainers and distributions can fix and ship kernel code.
OpenAI has not named a CVE list, so defenders cannot yet check which kernel flaws were involved.
Cost drives the comparison. A bill of $1,213,430 for 948 exploits is large for an individual but modest next to the budgets of well-funded organizations, the audience OpenAI describes with its phrase “well-resourced attacker.”
