OpenAI pointed a five-stage agentic pipeline, driven by an unreleased cyber model, at the Linux kernel in a controlled research environment — and the results show just how far automated vulnerability research has come. The pipeline identified 2,308 candidate findings, reproduced 1,724 vulnerabilities, and produced 948 working local privilege-escalation exploits. The run cost approximately $1.21 million in API credits, or about $1,279 per exploit.
The experiment is one of the most striking demonstrations yet of AI's ability to do sustained, multi-step security research at scale. It is also a case study in the economics of that capability: powerful, but not yet cheap.
The Linux Kernel: The Hardest Target in Software
The Linux kernel is the core of the Linux operating system, managing hardware, memory, processes, and security boundaries for everything from Android phones to cloud servers to supercomputers. It is also one of the most heavily reviewed codebases on earth, maintained by thousands of contributors and scrutinised by security researchers, fuzzing tools, and static analysers for decades.
That is what makes the numbers so striking. Finding vulnerabilities in the kernel is not like finding them in a small web application. The code is vast, the interactions between subsystems are subtle, and a candidate flaw must survive rigorous verification before it counts as a real vulnerability. A pipeline that can generate 2,308 candidate findings and confirm 1,724 of them as reproducible vulnerabilities is operating at a scale no human team could match.
What Is a Local Privilege-Escalation Exploit?
A local privilege-escalation exploit is a piece of code that lets an attacker who already has limited access to a system — say, an ordinary user account — gain higher privileges, typically full administrative or "root" access. Once an attacker has root, they can read any file, install persistent malware, disable security tools, and move laterally across a network.
Local privilege escalation matters because it is the second stage of most real-world attacks. The first stage gets an attacker a foothold; the second stage turns that foothold into control. A working exploit is not merely a theoretical description of a flaw — it is code that demonstrably breaks the security boundary, which is why producing 948 of them is a far stronger result than merely listing 948 suspected bugs.
The Five-Stage Agentic Pipeline
The pipeline OpenAI used had five stages, driven by an unreleased cyber model — a model not yet available to the public. An "agentic" pipeline differs from a single prompt to a chatbot: it is a system in which the model plans, acts, observes the results, and adjusts, chaining together many steps over a long run.
The staged design reflects how security research actually works. The first stages cast a wide net, generating candidate findings. Later stages filter and verify, reproducing vulnerabilities and discarding false positives. The drop from 2,308 candidates to 1,724 reproduced vulnerabilities to 948 working exploits shows that filtering in action: each stage demands more proof, and only the strongest findings survive to the end.
The Economics: $1.21 Million for 948 Exploits
The run cost about $1.21 million in API credits, working out to roughly $1,279 per working exploit. That figure is both impressive and sobering: $1,279 for a working kernel exploit is far less than the cost of a human security researcher spending weeks on the same target, yet $1.21 million for a single research run is a budget only the largest labs and best-funded organisations can afford.
For founders and investors, the cost curve is the story to watch. API prices for frontier models have fallen steeply over time, and agentic systems are becoming more efficient. If the cost per exploit falls by an order of magnitude — as it plausibly could over the next few years — automated vulnerability discovery moves from a showcase experiment to a routine capability, available to security teams of every size. The same economics, of course, apply to attackers.
Why a Controlled Research Environment Matters
OpenAI ran the experiment in a controlled research environment, using an unreleased model. That framing matters. The kernel vulnerabilities found in such a run would normally go through responsible disclosure — reported to the maintainers so they can be fixed before the details become public. The controlled setting is what separates security research from an attack: same techniques, different intent and process.
Why It Matters for Founders and Investors
For security founders, the experiment is a signal that the market for automated vulnerability research is about to expand dramatically. The tools to find flaws are becoming cheaper and more capable; the scarce resource becomes the ability to triage, prioritise, and fix findings inside real organisations.
For investors, the $1,279-per-exploit figure is a useful benchmark for the economics of AI-driven security research. Watch it fall. When the cost of finding a critical vulnerability drops below the cost of a human hour, the entire security market reorganises around automation — and the companies that reorganise first will win.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



