OpenAI’s AI Revealed 948 Linux Vulnerabilities. Now Comes the Hard Part

Artificial intelligence has demonstrated a new level of capability in cybersecurity. OpenAI says an automated system powered by an upcoming AI model identified and reproduced thousands of potential weaknesses in the Linux kernel, eventually producing working exploits for hundreds of distinct security flaws.

The research, published on November 4, 2026, reports that the system reproduced 1,724 vulnerabilities and generated 1,169 working local privilege-escalation exploits. After analysing related findings, OpenAI identified 948 distinct vulnerabilities.

The experiment cost $1.21 million in API credits and also produced proposed security patches for human review.

The findings matter because Linux supports much of the technology people and businesses rely on, including cloud infrastructure, web servers and Android devices. However, OpenAI has not established that all these products are affected, or that criminals have used the vulnerabilities in real-world attacks.

The research raises an important question for the technology industry: can software developers find and fix security weaknesses quickly enough as AI makes vulnerability discovery faster?

What Did OpenAI’s AI Actually Find?

OpenAI published its research under the title A Thousand Fewer Paths to Root.

The company developed an automated system to examine the Linux kernel, test suspected security weaknesses and assess whether they could be exploited.

The research took place in mid-October 2026, although OpenAI published the results on November 4.

The system began by identifying 5,034 areas of the Linux kernel for closer examination. These covered approximately 2.8 million lines of code, around 7% of the codebase.

It then produced 2,308 potential vulnerability reports. These were not all confirmed security flaws. Each finding needed further testing to establish whether it represented a genuine weakness.

The system reproduced 1,724 of those findings in isolated virtual machines.

It then moved to the next stage: testing whether the vulnerabilities could allow a user with limited system access to gain greater control.

According to OpenAI’s detailed results, the agents generated 1,169 working local privilege-escalation exploits. After analysing related findings and removing duplicates or variants, the company counted 948 distinct vulnerabilities.

The system also produced 291 proof-of-concept programs demonstrating information leaks.

These results explain why the 948 figure should not be confused with the total number of exploit programs or the number of potential findings.

What Does Gaining Root Access Mean?

The vulnerabilities involved a security weakness known as local privilege escalation.

Linux systems use different permission levels to control what users and applications can do. An ordinary account has limited permissions, while the root account has extensive administrative control.

A local privilege-escalation vulnerability can allow someone who already has limited access to a system to gain those higher permissions.

For example, an attacker who has obtained access to a restricted account might exploit such a weakness to gain control over protected system resources.

That could increase the damage they can cause, depending on the system, the vulnerability and the protections in place.

However, local privilege escalation is not the same as remotely hacking a device without any prior access.

OpenAI’s research demonstrated vulnerabilities that could allow an unprivileged local user to obtain root access. It did not demonstrate that someone could take control of an arbitrary Linux device simply by visiting a website or opening a message.

That distinction is important when assessing the potential risk to ordinary users.

How Did the AI Discover and Test the Vulnerabilities?

The most significant part of the research is the way OpenAI combined several tasks into one automated process.

Rather than simply asking an AI model to inspect code and report suspicious sections, the company developed a pipeline with several stages.

First, specialised agents selected areas of the Linux kernel that were relevant to security and could be reached by users with limited permissions.

Other agents examined the selected code and identified possible weaknesses.

The system then attempted to reproduce each suspected vulnerability inside an isolated virtual machine. This helped determine whether the reported issue could actually be triggered.

Once a weakness had been reproduced, further agents assessed whether it could be exploited to gain higher privileges or expose information.

The system also generated proposed patches to address the problems.

Finally, human security researchers reviewed the results, including the exploits and proposed fixes, before reporting vulnerabilities to maintainers.

This process matters because identifying suspicious code is not enough to prove that a security vulnerability exists. Testing helps establish whether a weakness is real and what an attacker might be able to achieve.

OpenAI described the process as an automated vulnerability discovery, exploitation and patch-generation pipeline.

Why Did OpenAI Spend $1.21 Million?

The full research run cost $1,213,430 in API credits, according to OpenAI.

The company calculated an average cost of approximately $1,279 for each of the 948 distinct local privilege-escalation vulnerabilities.

That figure covers the particular research experiment. It should not be treated as the standard cost of discovering a Linux vulnerability or as a price that every organisation would pay to use similar technology.

OpenAI also compared the cost with historical rewards offered by the Pwn2Own and TyphoonPWN security competitions.

Using historical reward amounts of $20,000 and $70,000 for relevant Linux kernel privilege-escalation exploits, OpenAI calculated a hypothetical value of approximately $18.9 million to $66.3 million.

This was an illustrative comparison, not money earned or a confirmed valuation of the vulnerabilities. There is no indication that all 948 findings qualified for those rewards.

The comparison nevertheless highlights why AI-assisted vulnerability research could become an important tool for organisations responsible for securing complex software.

Did the AI Discover New Hacking Techniques?

OpenAI reported another notable finding.

Researchers examined a sample of 166 exploits generated by the system. They concluded that, to the best of their knowledge, the agents had used at least 10 exploitation techniques that were new to the researchers and the outside experts they consulted.

These techniques involved different approaches to working with kernel memory and obtaining higher privileges.

OpenAI has not published their full technical details, citing the security risks of disclosing new exploitation methods.

It is important to understand the scope of this finding. The company identified the techniques through an analysis of a sample of 166 exploits. It did not claim that every exploit used a completely new technique.

The result suggests that advanced AI systems can do more than recognise previously documented vulnerabilities. Under certain conditions, they may also develop new ways to demonstrate how software weaknesses can be exploited.

Are the Linux Vulnerabilities Already Fixed?

Not all the findings were resolved when OpenAI completed its research.

The company said it worked with Linux maintainers and Patch the Planet partners to report vulnerabilities and provide proposed patches.

It also shared a snapshot of the repair activity during the research period.

Within the 10-day period covered by its report:

• 61 local privilege-escalation vulnerabilities had been patched upstream.
• 43 had staging patches.
• Another 299 were being publicly discussed or reported.

These figures describe the situation observed during the research period. They should not be presented as a live count of vulnerabilities that remain unpatched today.

There is also an important difference between an upstream Linux fix and an update reaching an individual device.

A patch may first be reviewed and accepted into the Linux kernel development process. Linux distributions and device manufacturers then need to incorporate relevant fixes into their own software releases.

This means a fix being available upstream does not automatically mean every affected computer, server or Android device has received it.

OpenAI did not publish a complete list of affected kernel versions in the announcement. Readers therefore cannot use the 948 figure alone to determine whether their particular device is affected.

Could Android Phones and Other Devices Be Affected?

Linux is not limited to computers running a traditional Linux desktop.

The kernel is used in Android, cloud infrastructure, servers and many other technology systems. This gives the research significance beyond the developers who work directly on Linux.

Android, for example, uses a Linux-based kernel. However, that does not mean the research has established that Android phones contain all 948 vulnerabilities.

The actual impact depends on which vulnerabilities affect a particular kernel version, whether the relevant code is present and reachable, and what security protections the device has.

The same applies to cloud services and other Linux-based systems.

OpenAI’s announcement does not identify specific affected Android models, Linux distributions, cloud providers or device manufacturers.

For that reason, it would be misleading to tell readers that their phones or personal information are currently exposed because of this research.

The more accurate conclusion is that the discovery concerns a widely used software component, making the work of identifying and fixing relevant weaknesses important across the technology industry.

What Is the UK National Cyber Security Centre Saying About AI Risks?

OpenAI’s research also comes at a time when government cybersecurity agencies are examining the wider impact of AI on cyber threats.

In its separate assessment, Artificial Intelligence and Cyber Intrusion, the UK’s National Cyber Security Centre (NCSC) examined how AI could change the cyber threat landscape through to 2028.

The assessment warns that AI tools could help attackers identify and exploit software weaknesses more efficiently. It also explains that defensive AI tools could help organisations find vulnerabilities and apply fixes more quickly.

The NCSC stresses that organisations will need to improve their patching and testing processes to keep pace with these developments.

This is a broader assessment of future cybersecurity risks, not a specific investigation into OpenAI’s Linux research.

Together, the two developments highlight an important issue: AI can help both sides of cybersecurity, but organisations still need effective security controls, testing and timely software updates.

What Should Linux Users and Businesses Do?

OpenAI’s findings do not establish that ordinary users need to take emergency action or replace their devices.

However, the research reinforces the importance of basic security practices.

For businesses and system administrators:

• Install security updates released by your Linux distribution or software provider.
• Monitor security advisories relevant to the systems you operate.
• Review access permissions and avoid giving ordinary accounts unnecessary administrative privileges.
• Use isolation and sandboxing to limit the impact of a compromised process.
• Test security patches before deploying them across important systems.

For ordinary users, the most practical step is to keep their operating systems and devices updated through official update channels.

Android users should install security updates provided by their device manufacturer when available.

These steps are sensible security practices. They are not evidence that a particular user’s device is affected by OpenAI’s findings.

What This Discovery Really Means

OpenAI’s research demonstrates how much vulnerability research could change as AI systems become more capable.

The system did not simply identify suspicious lines of code. It tested potential weaknesses, reproduced vulnerabilities, generated working exploits and prepared proposed fixes for human review.

That could give software developers a faster way to investigate security problems and address them before criminals can take advantage of them.

There is also a serious challenge. If similar capabilities become more widely available, attackers could use AI to investigate software weaknesses more quickly too.

But OpenAI’s research does not establish that criminals have used these particular Linux vulnerabilities in attacks, or that the upcoming model is already publicly available.

The important question is how quickly software developers, maintainers and security teams can adapt to a world in which finding and testing vulnerabilities may become increasingly automated.

For users, the immediate lesson is straightforward: keep software updated, use supported devices and follow security guidance from the organisations responsible for maintaining your systems.

EDITORIAL TRANSPARENCY

How we checked this

We reviewed OpenAI’s original research, checked its vulnerability and exploit figures against the published methodology, and consulted The Hacker News and the UK’s National Cyber Security Centre for additional context. We have distinguished demonstrated findings from unconfirmed risks and noted where information, including affected kernel versions, remains undisclosed.