OpenAI AI pipeline finds 948 Linux kernel exploits in one run

A massive leap in speed

OpenAI recently demonstrated the power of its cyber-focused AI pipeline by running it against the Linux kernel. The system identified 948 working privilege escalation exploits, which are flaws that allow a standard user to gain root access. Each of these findings came with a candidate patch already attached to the result.

This achievement represents a significant jump in speed compared to traditional methods. Usually, human researchers and fuzzing tools surface only a few hundred serious flaws across an entire year of work. In contrast, OpenAI's system produced roughly double that amount in a single run.

The results were published in a blog post titled "A thousand fewer paths to root." As Startup Fortune reported, the numbers show that the bottleneck in AI-assisted vulnerability research is shrinking rapidly. The company is moving from staggered, months-long research cycles to nearly instantaneous results.

How the pipeline functions

The process began by sifting through 2,308 candidate findings within the Linux kernel. The AI pipeline successfully reproduced 1,724 of these as real, working vulnerabilities. From that pool, it distilled the final batch of 948 validated local privilege escalation exploits.

The system did more than just find known bug classes faster. It also surfaced at least 10 previously unknown Linux exploitation techniques. This means the model is capable of discovering new ways to turn kernel bugs into working exploits that researchers had not documented before.

While the AI does the heavy lifting of finding and drafting patches, human oversight remains a core component. OpenAI works with the Patch the Planet program, where engineers from Trail of Bits review every finding. This ensures that security-critical infrastructure is not patched based on AI output alone without human verification.

The economics of security

The cost of this specific run was approximately $1.2 million in API spend. When broken down, this works out to roughly $1,265 per vulnerability. This price point is notable because it delivers a complete exploit-and-patch pair in one sitting rather than over a year of human labor.

For companies running security operations, these numbers change the math of automated auditing. The cost of the AI run is comparable to, or even less than, the annual salary of a single senior security engineer or a dedicated red team. This makes automated scanning a more viable budget item for startups and large enterprises alike.

The shift is already creating a new market shape for security tools. Startups building automated pentesting and code scanning now face a high bar for competition. They must compete against a track record where AI can handle the bulk of the discovery work at a predictable price point.

Evolution of the tool

This result is part of a broader pattern OpenAI has been developing over several months. In May 2025, an independent researcher used the o3 model to find a remote zero-day in the Linux kernel's SMB implementation. That patch was successfully shipped shortly after the discovery.

In June 2026, OpenAI formally launched the Patch the Planet initiative. This program pairs a cyber-specialized model with human engineers to review findings before they reach project maintainers. At its launch, the same approach produced a much smaller number of exploits than the recent run.

The jump from 24 local privilege escalation proofs of concept at launch to 948 in just four months is a different order of magnitude. It suggests that the underlying models are becoming significantly more proficient at navigating complex kernel code. This progression indicates that the technology is moving toward a production-ready roadmap.

What comes next for the industry

The primary question for the industry is no longer whether AI can find kernel bugs at scale. The results clearly show that it can. The focus is shifting toward who will have access to these pipelines and how they will be licensed or built in-house.

OpenAI is also continuing to develop other autonomous agents, such as Aardvark. That agent was designed for autonomous vulnerability hunting and has already found 92% of known and seeded vulnerabilities in testing. These tools are building a portfolio of capabilities that go beyond simple bug finding.

As these tools become more common, the speed of patching will likely increase. However, the necessity of human review remains a constant. Security-critical systems require a level of certainty that currently necessitates a human engineer to sign off on every AI-generated fix before it goes live.

The path to production

OpenAI's work with the Linux kernel maintainers highlights a collaborative approach to security. By working through the Patch the Planet program, the company aims to push findings toward real fixes rather than releasing exploit code into the wild. This responsible disclosure model is essential for maintaining trust in the security community.

The transition from research papers to product roadmaps is accelerating. As the costs of finding vulnerabilities drop, the focus will turn to the volume of patches that maintainers can actually process. The pipeline may find thousands of bugs, but the human bottleneck of applying those patches remains a factor.

For the average user, this means that the underlying software powering their devices may become more secure more quickly. While the average person won't see the $1.2 million spend, they will benefit from a faster cycle of vulnerability discovery and remediation in the core systems of the internet.

The impact on developers

Developers and security teams will see a shift in how they prioritize their work. Instead of spending months on manual fuzzing and bug hunting, teams may spend more time reviewing high-quality AI-generated candidates. This allows human expertise to be applied where it is most needed: on the most complex and nuanced flaws.

The availability of candidate patches for every exploit found is a major advantage for rapid response. It allows security teams to see the proposed fix immediately alongside the vulnerability. This transparency helps in assessing the risk and the viability of a patch before it is deployed to production environments.

As these models become more integrated into the development lifecycle, the distinction between 'manual' and 'automated' security may blur. The goal is a hybrid system where AI provides the scale and speed, while humans provide the final verification and strategic oversight for the world's most critical software infrastructure.

Scaling the security pipeline

The success of the Linux kernel run provides a blueprint for other types of software auditing. While the kernel is one of the most complex pieces of software, the principles of using a cyber-specialized model to find and patch flaws can be applied to various other systems. This could eventually lead to automated auditing for entire software stacks.

The scalability of this approach is what makes it a significant milestone. By automating the reproduction of vulnerabilities and the generation of patches, OpenAI has removed several manual steps from the security research pipeline. This allows for a much higher volume of work to be completed in a fraction of the time.

We can expect to see more companies adopting similar pipelines to secure their own proprietary codebases. The ability to find and fix bugs at a cost of roughly $1,265 per vulnerability provides a compelling economic argument for enterprises to invest in AI-driven security tools over traditional manual methods.

Future of autonomous agents

The trajectory of OpenAI's tools suggests a move toward fully autonomous security agents. Aardvark is a prime example of this, already earning 10 CVEs through responsible disclosure. These agents are designed to navigate the complexities of open-source bugs and find vulnerabilities without constant human intervention.

As these agents become more sophisticated, they will likely be able to handle more diverse tasks. This includes not just finding bugs, but also performing automated penetration testing and complex code analysis. The goal is to create a system that can proactively defend against threats before they can be exploited by malicious actors.

The ultimate goal of these initiatives is to make the software world more resilient. By leveraging the speed of AI to find and patch flaws, the industry can move toward a state where vulnerabilities are addressed as soon as they are discovered, significantly narrowing the window of opportunity for cyberattacks.

The role of human experts

Despite the impressive scale of the AI's findings, the human element remains indispensable. Trail of Bits engineers are the ones who ultimately sign off on the findings. This human-in-the-loop model is a necessary safeguard against the potential for AI to generate incorrect or harmful patches.

Human experts provide the context and nuance that AI models may still lack. They understand the broader implications of a patch, including how it might affect performance or compatibility with other systems. This expertise ensures that the fixes provided by the AI are safe for wide-scale deployment in critical infrastructure.

The future of security research will likely be a partnership between human intelligence and machine scale. AI will act as a force multiplier, allowing a smaller number of experts to oversee a much larger volume of security work. This collaboration is what will drive the next generation of software security and vulnerability management.

Economic shifts in cybersecurity

The economic implications of these results are profound for the cybersecurity industry. When a single run can produce nearly 1,000 validated exploits and patches, it disrupts the traditional business models of security research firms. Companies that rely on manual discovery will need to adapt to a world where AI can do the heavy lifting at a lower cost.

This shift will likely lead to a decrease in the cost of security for end-users and enterprises. As the price per vulnerability drops, more software can be audited more frequently. This could lead to a higher standard of security across the entire software ecosystem, from core kernels to consumer applications.

For startups in the AI security space, the challenge will be to provide unique value beyond just finding bugs. They will need to offer specialized knowledge, better integration with existing workflows, or more sophisticated human-led verification services. The market is moving toward a model where AI is the engine, but human oversight remains the steering wheel.

A new standard for auditing

OpenAI's results set a new benchmark for what is possible in automated code auditing. By identifying 948 exploits in a single run, they have demonstrated that AI can handle the most complex parts of the Linux kernel. This paves the way for similar audits of other major operating systems and critical software components.

The ability to generate candidate patches alongside each exploit is a key differentiator. It moves the needle from just identifying problems to providing immediate solutions. This end-to-end capability is what will make AI-assisted security a standard part of the software development lifecycle for many organizations.

As these tools become more accessible, the speed of the 'patch cycle' will continue to accelerate. The time between a vulnerability being discovered and a fix being deployed will shrink, making it harder for attackers to exploit known flaws. This is a significant step toward a more secure and resilient digital world for everyone.

The path forward for Patch the Planet

The Patch the Planet program serves as a model for how AI can be used responsibly in cybersecurity. By partnering with established security firms and open-source maintainers, OpenAI ensures that its AI-generated findings are handled correctly. This collaborative framework is essential for building trust in AI-driven security tools.

The program's success with the Linux kernel shows that AI can be a valuable partner for maintainers. It provides them with a steady stream of high-quality, pre-validated findings that can be reviewed and integrated. This helps to alleviate the burden on maintainers who are often overwhelmed by the volume of security reports they receive.

As the program matures, we can expect to see it expand to other major open-source projects. The goal is to create a global infrastructure for AI-assisted security, where models are used to proactively find and fix flaws in the software that powers our modern world. This is the next frontier of the AI revolution in cybersecurity.

Summary of the kernel run

To summarize, OpenAI's AI pipeline has achieved a remarkable feat by identifying 948 privilege escalation exploits in the Linux kernel. The process was efficient, costing $1.2 million in API spend and producing results in a single run that would typically take humans a year to achieve. Each exploit came with a candidate patch, ready for human review.

The run also highlighted the model's ability to discover new exploitation techniques, showing that it is not just replicating known patterns. By identifying 10 previously unknown techniques, the AI demonstrated a sophisticated understanding of kernel mechanics. This capability is a major step forward for automated vulnerability research.

While human review remains a critical final step, the economics of this run are clear. The cost per vulnerability is low enough to make automated auditing a viable strategy for many organizations. This marks a significant shift in how we think about, and execute, software security in the age of AI.

The impact on the security market

The security market is already reacting to these results. Startups and established firms alike are looking at how to integrate these types of AI pipelines into their services. The ability to deliver high-volume, high-quality results quickly is a major competitive advantage in the cybersecurity space.

We can expect to see more products that offer 'AI-assisted' security auditing as a standard feature. These tools will likely be marketed to enterprises as a way to scale their security efforts without linearly increasing their headcount. The goal is to provide a way to secure more code with fewer human resources.

As these tools become more common, the focus will shift toward the quality of the human review process. Companies will need to ensure they have the right experts to verify the AI's work. The human-in-the-loop model will be the standard for ensuring that AI-generated patches are safe and effective for production use.

The future of AI in security

The future of AI in security is one of massive scale and rapid iteration. OpenAI's kernel run is a clear indicator of this trend. As models become more capable and costs continue to fall, the ability to audit and secure software will become increasingly automated.

This will lead to a more proactive security posture, where vulnerabilities are found and patched before they can be exploited. It will also allow human experts to focus on the most complex and high-impact security challenges. The AI will handle the volume, while humans provide the strategy and final verification.

This evolution is already underway. From the o3 model's early successes to the Patch the Planet initiative, OpenAI is building a framework for a new era of cybersecurity. In this era, AI is not just a tool for researchers, but a fundamental component of the infrastructure that keeps our digital world safe.

Final thoughts on the kernel run

The results from the Linux kernel run are a landmark moment for AI in cybersecurity. By identifying 948 exploits and patches in a single run, OpenAI has shown that AI can perform complex security research at an unprecedented scale. This is a major step toward making software more secure for everyone.

The economics of the run also provide a clear path forward for the industry. The cost per vulnerability is low enough to make automated auditing a practical reality for many organizations. This will lead to a shift in how companies approach security, moving toward a more automated and proactive model.

While human oversight remains essential, the role of AI as a force multiplier is undeniable. As these tools become more integrated into the security landscape, they will help to create a more resilient and secure digital world. The journey toward AI-driven security is well underway, and the results are already showing their impact.

The role of open source

The Linux kernel is a cornerstone of the open-source world, and its security is vital for the entire internet. By focusing on the kernel, OpenAI is addressing one of the most critical pieces of software infrastructure. This work benefits everyone who relies on Linux, from server administrators to everyday users.

The collaborative nature of the Patch the Planet program also highlights the importance of open-source cooperation. By working with maintainers and security firms, OpenAI is ensuring that its AI-driven findings are used to improve the software for the common good. This is a model that can be replicated across many other open-source projects.

As more open-source projects adopt AI-assisted security, the overall security of the software ecosystem will improve. This will lead to a faster cycle of vulnerability discovery and remediation, making it harder for attackers to find and exploit flaws. It is a significant step toward a more secure and resilient digital world for all.

The next steps for OpenAI