OpenAI has demonstrated that large models can automate broad swaths of vulnerability research by building an autonomous pipeline that discovered, validated and produced candidate patches for Linux kernel vulnerabilities at scale. The prototype scanned selected areas of the kernel, generated proofs of concept and exploit variants, and delivered candidate fixes and human-review artifacts to maintainers.
The system is organized into five stages: vulnerability discovery, dynamic validation, exploitability assessment, patch generation and human review. At the core of the prototype was a forthcoming cyber-focused model used to power autonomous agents that performed targeted auditing, crafted minimal crashing proofs, attempted exploit development, and proposed source-code changes. For each confirmed finding the pipeline bundled a root-cause analysis, a minimal crashing proof-of-concept, a candidate patch and a full exploit to aid human reviewers and upstream maintainers.
Discovery began with Autonomous Scout agents that selected audit targets by analyzing reachability and security relevance. That process identified 5,034 audit targets—about 2.8 million lines of code, roughly 7% of the Linux kernel. Analyst agents then produced 2,308 deduplicated candidate findings within 12 hours, with the analyst scanning step accounting for the largest early-stage cost in API credits.
Triage agents moved candidate findings into dynamic validation using isolated virtual machines configured with Kernel Address Sanitizer (KASAN). Within 30 hours the system dynamically reproduced 1,724 vulnerabilities, representing about 74.7% of the initial candidates. The validation stage had a published API cost of approximately $73,330 in credits.
A subsequent screening step categorized validated issues into 313 information-leak candidates and 1,238 local privilege-escalation (LPE) candidates. Exploit-development agents then tested candidates in VMs configured with standard kernel protections—SMEP, SMAP, KASLR and hardened usercopy. From that phase the agents produced 291 information-leak proofs of concept and 1,169 working LPE exploits; variant analysis reduced the LPE set to 948 distinct exploit variants linked to the underlying vulnerabilities. All outputs were reviewed by human researchers before reporting to maintainers.
OpenAI published detailed API cost figures for each pipeline stage to support reproducibility and defensive planning. Key reported costs include $3,426 for scouting, $316,549 for analyst scanning, $73,330 for validation, $44,194 for initial screening, and $775,929 for the exploitation stage (split into $84,929 for information leaks and $691,000 for LPE candidates). Successful exploits averaged about $359 in credits to develop, while a smaller set of harder-to-exploit cases consumed much larger amounts.
To contextualize economics, the report compared pipeline costs to historical bounty levels from events such as Pwn2Own and TyphoonPWN, noting that applying past reward levels to the 948 exploits would yield hypothetical valuations many times the pipeline cost. In a targeted run against Google’s kernelCTF environment the prototype obtained 25 information leaks and 21 LPEs at a campaign cost of roughly $32,936 over 112 hours, figures OpenAI presents as benchmarks for defensive planning.
Sampling 166 agent-generated exploits, OpenAI’s researchers identified 10 exploitation techniques they judged as previously unseen; technical details were shared with third-party experts but omitted from public discussion for safety reasons. The October 2026 campaign also surfaced a QEMU TCG vulnerability in the test infrastructure: that issue had already been fixed upstream but remained present in the release image used for testing, underlining the researchers’ point that research infrastructure itself must be secured when running automated offensive workflows.
For each validated vulnerability the pipeline proposed candidate patches. Iterative prompting and additional checks improved patch quality, reducing extraneous changes and aligning fixes with local coding practices. OpenAI reports hundreds of critical vulnerabilities and patches have been submitted to maintainers, and that the team is working with Patch the Planet partners and the Linux community to upstream fixes responsibly.
Maintainer response in the reported campaign was prompt: within a 10-day window 61 LPEs had been patched upstream, 43 had staging patches and 299 were under public reporting and discussion. The authors frame the work as evidence that frontier models can automate many stages of vulnerability research—discovery, dynamic validation, exploitability assessment and patch generation—while providing cost and capability data to help defenders make exploitation more expensive for attackers.
OpenAI documents the work and shares findings with maintainers to support increased software security. By publishing methods, costs and outcomes, the project aims to give defenders actionable information about the risks posed by automated workflows and to encourage proactive hardening where it will most effectively reduce the attack surface for Linux kernel vulnerabilities.
Source: Read the original source
Leave a Reply