跳到正文
DEV Community· News·· 18 天前AI 評分74

OpenAI 詳述自主漏洞發現流程:耗用 $1,213,430 產出 948 個 Linux 核心本地權限提升漏洞利用

OpenAI details autonomous pipeline that found 948 Linux privilege-escalation flaws in one 1.2M dollar run

AI 導讀

OpenAI 公佈一條自主漏洞發現流程,在一次自動化 Linux 核心安全審查中產出 948 個可運作的本地權限提升漏洞利用。流程提出 2,308 個候選核心問題,經動態驗證確認 1,724 個為真實漏洞;本次運行耗用 $1,213,430 API 額度,平均每個可運作漏洞利用約 $1,279。這次工作在受控隔離環境進行,輸出須經人工審核;候選補丁仍要通過 Linux 核心維護者的正常上游審查。

正文

Infographic cover

📅 November 4, 2026 • openai.com

OpenAI has published details of an autonomous vulnerability-discovery pipeline that completed a large-scale security review of the Linux kernel in a single automated run. The system is powered by an upcoming cyber model that the company has not yet named. It moves through four defined stages: discovery, dynamic validation, exploitability assessment and patch generation. Each stage hands its output to the next, so a flaw surfaced by the discovery module is confirmed, classified, assessed for exploitability and paired with a fix without human involvement until the final review gate. In practice that means every reported finding arrives with its full validation and patching history attached.

The run took place in mid-October against a stock desktop Linux distribution inside a controlled, isolated research environment. Per OpenAI, this was a laboratory exercise rather than a response to attacks in the wild, and no live system was breached at any point. All of this happened against the packaged distribution as shipped, not a specially prepared target. The pipeline surfaced 2,308 candidate kernel findings. Dynamic validation reproduced 1,724 of them as genuine vulnerabilities. The exploitability assessment produced 291 proofs of concept for information leaks. The final stage delivered 948 working local privilege-escalation exploits, and each of them ships with a candidate patch, a root-cause analysis and a minimal crashing proof of concept.

The economics of the run are as notable as the technical results. Per OpenAI, the full campaign cost $1,213,430 in API credits, which works out to roughly $1,279 per working exploit. Those two figures give security teams something they rarely have in writing: a budget line that converts into a known number of reproduced, documented and patchable kernel flaws, each documented well enough for a maintainer to act on.

Human researchers stay in the loop at the end of the pipeline. Per OpenAI, every output is reviewed before any report leaves the lab, so nothing reaches maintainers or partners without human sign-off. Reporting started with the most critical vulnerabilities, in line with Linux kernel community guidance, and disclosure is coordinated through Patch the Planet partners.

The work expands the June 2026 Patch the Planet effort, in which GPT-5.5 Cyber produced 24 Linux kernel local privilege-escalation exploits. The new run delivers roughly forty times that output. All patches produced by the pipeline remain candidates until they pass through normal upstream maintainer review, where the usual kernel patch process takes over.

IMPACT: Security research is turning into a systematic, budgeted pipeline — generate, reproduce, prove, patch, and only then a human gate. For kernel maintainers that means rising patch velocity on long-tail flaws that once waited years for attention, with candidate patches arriving through the normal upstream review process rather than around it.

Sources: openai.com

Confirmations could not be found at this time.

Key figures

IMPACT checklist


Sources:


Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

AI Automation

來源:DEV Community · dev.to