In a demonstration of the accelerating power of artificial intelligence in cybersecurity, OpenAI has revealed that an autonomous pipeline powered by an unreleased, specialized cyber model successfully engineered root-level exploits for 948 unique Linux kernel vulnerabilities. This massive undertaking, conducted during a single automated run, cost $1.2 million in API credits, marking a significant milestone in how machine learning can be leveraged for both offensive and defensive security operations.
The ambitious project, which took place in mid-October, focused on the latest code residing in the Linux kernel’s master branch. The AI agents were tasked with performing a deep-dive analysis of the kernel’s architecture. Initially, the agents successfully narrowed their focus to security-relevant sections of the code that are reachable by low-privileged users—a critical vector for privilege escalation attacks. Following this narrowing process, the system flagged 2,308 potential bugs and, through its autonomous verification process, confirmed the existence of 1,724 of them.
To ensure the exploits were not merely theoretical, the agents tested their findings on virtual machines running an unmodified desktop Linux distribution. These environments were not "stripped down" for testing; instead, they included standard modern security protections, such as Kernel Address Space Layout Randomization (KASLR), Supervisor Mode Execution Protection (SMEP), and Supervisor Mode Access Prevention (SMAP). Despite these robust defenses, the agents produced 1,169 working local privilege escalation (LPE) exploits, alongside 291 proofs of concept designed to demonstrate information leaks. OpenAI clarified that the final count of 948 unique vulnerabilities reflects the number of distinct patches generated to address the bugs identified behind those exploits. Before any findings or patches were submitted to the kernel maintainers, they underwent a rigorous review process by OpenAI’s own internal security researchers, who vetted the exploits, the proposed fixes, and the severity levels assigned by the model.
Breakthroughs in Exploitation Techniques
Perhaps the most startling aspect of the report is the sophistication of the vulnerabilities discovered. OpenAI stated that, to the best of their knowledge, the agents in their pipeline developed at least 10 novel Linux exploitation techniques. These methods were sophisticated enough that even seasoned kernel exploitation experts outside of OpenAI were unfamiliar with them. Given the potential for misuse, the company has opted to keep the technical details of these novel techniques private, citing safety concerns and the need to prevent the proliferation of such potent attack vectors.
The financial breakdown of the project provides a glimpse into the economics of AI-driven vulnerability research. The total bill for the October scan reached $1,213,430. When spread across the 948 exploits discovered, the cost per exploit averages roughly $1,279. This represents a significant increase in capability compared to an earlier, smaller-scale scan conducted in late July, which was powered by a predecessor model, GPT-5.6 Sol. That previous iteration yielded 166 unique LPEs at a cost of $513,690.
OpenAI emphasized the immense value of this research by comparing the costs to current market rates for security flaws. In the world of professional bug bounties, platforms like Pwn2Own and TyphoonPWN have historically paid between $20,000 and $70,000 for a single Linux kernel privilege escalation exploit. By this metric, the 948 bugs discovered by the AI would be worth between $18.9 million and $66.3 million on the open market. In other words, the autonomous scan was between 16 and 55 times more cost-effective than traditional human-led vulnerability research.
The effectiveness of this new model is further underscored by the fact that the October scan uncovered more than five times as many new LPEs as the July scan, despite the kernel having undergone three months of intensive hardening in the interim. OpenAI also tested the pipeline against an adapted version of Google’s kernelCTF—a hardened environment specifically designed to have a smaller attack surface. Even in this restricted environment, the system delivered 21 LPE exploits over the course of 112 hours of compute time, at a cost of approximately $33,000 in API credits.
Shaping the Future of Patching and Defense
In the wake of these findings, OpenAI has been working closely with the kernel community to ensure the security of the broader ecosystem. Following guidance from maintainers, the company has already submitted hundreds of reports and corresponding patches for the most critical flaws. The remaining bugs are slated to be shared in stages, with the company coordinating through its "Patch the Planet" initiative to ensure that maintainers are not overwhelmed and that the most critical systems are prioritized.
The impact of this effort has already been felt in the open-source community. OpenAI noted that within just 10 days of the scan’s commencement, they observed that over 100 Linux LPEs had either been patched or had staging patches in development by other contributors, with approximately 300 additional flaws currently under public discussion.
However, the company’s findings serve as a stark warning to organizations and system administrators. OpenAI’s assessment suggests that the threat landscape is shifting rapidly. They have cautioned defenders that, in the near to medium term, any low-privileged account on an operating system that is not fully up-to-date with the latest patches should be considered functionally equivalent to root access for a well-resourced threat actor. To mitigate these risks, OpenAI strongly advocates for a "defense-in-depth" approach, urging organizations to move beyond relying solely on patching and to implement robust, layered defenses, including advanced sandboxing and isolation technologies such as Firecracker and gVisor.
OpenAI Raises Bug Bounty Rewards
Recognizing that autonomous systems cannot yet catch every flaw, OpenAI is simultaneously bolstering its human-led security efforts. On November 2, the company announced a significant increase in payouts for its Security Bug Bounty program. The objective is to incentivize outside security researchers to continue hunting for vulnerabilities that might elude even the most advanced AI testing pipelines.
Under the new structure, the midpoint of the payout ranges for medium, high, and critical-severity bugs has increased more than four times on average. A researcher who identifies and reports a critical flaw in core products such as ChatGPT, Codex, or the API platform can now receive a payout of up to $30,000. While the cap for exceptional critical bugs remains at $100,000, new bonus structures mean that a single high-quality, high-impact report could potentially result in a total payout of $172,000.
In addition to base compensation, OpenAI is introducing a five-tier loyalty program for researchers who consistently contribute valid bug reports, along with extra multipliers for vulnerabilities discovered during priority campaigns. By combining the massive, high-speed output of its autonomous pipelines with a heavily incentivized, human-led bounty program, OpenAI is positioning itself at the center of a new, hyper-accelerated era of software security. This dual-pronged strategy highlights a broader industry trend where the integration of AI into the security lifecycle is no longer a luxury but a fundamental necessity for maintaining the integrity of critical infrastructure.






