The company said on Wednesday that its experimental pipeline, powered by an unreleased cyber-focused model, reproduced 1,724 vulnerabilities from 2,308 candidate findings during a mid-October exercise. It also produced proofs of concept for 291 information disclosure flaws and proposed patches for human review.
The operation consumed $1,213,430 in application programming interface credits, equivalent to approximately $1,279 for each distinct privilege escalation exploit. OpenAI described the expenditure as the cost of the research run, rather than the expense of deploying approved fixes across affected systems.
Local privilege escalation vulnerabilities allow someone who already has limited access to a computer to acquire higher permissions, potentially including root privileges that confer control over the operating system. The findings do not establish that attackers remotely compromised 948 machines or that the flaws have been exploited outside the laboratory.
OpenAI tested its exploits against virtual machines running an ordinary desktop Linux installation rather than an intentionally weakened operating system. The research targeted the latest Linux kernel development branch available when the campaign began, although the company did not identify the distribution or name the forthcoming model.
The automated workflow selected areas of kernel code for inspection, searched for possible defects and attempted to reproduce them in isolated environments. Specialist agents then developed exploit demonstrations and candidate corrections. Human security researchers subsequently assessed the evidence, severity and proposed fixes before communicating findings to maintainers.
The company said it was prioritising critical vulnerabilities in consultation with the Linux kernel community and working with partners under its Patch the Planet initiative to submit additional reports. Producing a candidate patch does not mean the change has been accepted into the kernel or delivered to users through distribution updates.
The pace of maintenance also complicated the exercise. By the end of the ten-day scanning and review period, 61 privilege escalation vulnerabilities had already been patched upstream, 43 had patches awaiting integration and another 299 were being reported or publicly discussed. Those figures included work by other contributors, rather than fixes attributable exclusively to OpenAI.
The results substantially exceeded an earlier experiment conducted in late July using GPT-5.6 Sol. That campaign identified and exploited 166 distinct privilege escalation vulnerabilities at an API cost of $513,690. OpenAI said the October model discovered more than five times as many previously unknown exploitable weaknesses despite intervening kernel hardening.
Researchers also examined a sample of 166 exploits and identified at least ten techniques they considered novel. OpenAI said external kernel exploitation specialists consulted by the company had not encountered those approaches. It withheld detailed descriptions, citing security concerns about making the techniques easier to reproduce.
A separate assessment used an adapted version of Google’s kernelCTF testing environment, which incorporates additional restrictions intended to make exploitation harder. The pipeline generated 21 privilege escalation exploits and 25 information leak demonstrations during 112 hours of testing, consuming approximately $32,936 in API credits.
OpenAI compared those results with historical rewards offered by vulnerability competitions, but the projected financial value was hypothetical. The company did not establish that the exploits would qualify for payments or that any programme had accepted the findings. Its reported costs likewise do not represent an independently audited accounting of every research expense.
The disclosure raises practical questions for organisations operating shared Linux infrastructure, where a compromised low-privilege account can become a route to wider system control. OpenAI emphasised prompt security updates and additional isolation mechanisms to reduce damage when an attacker succeeds in exploiting a kernel weakness.
OpenAI said the system examined thousands of selected audit targets rather than every possible kernel execution path. Its findings therefore should not be interpreted as a complete inventory of Linux weaknesses. The research also illustrates a disclosure challenge: automatically generated exploits can accumulate faster than volunteer maintainers can investigate, reproduce and safely integrate corrections, particularly when fixes must be checked for regressions and compatibility with existing code.
The company also said its automated assessments must be repeated against updated code to distinguish unresolved problems from vulnerabilities already corrected. Such checks can help maintainers avoid duplicate reports while focusing limited review capacity on flaws that remain exploitable.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
