Jemiboo Magazine

Linux kernel vulnerabilities: OpenAI agents find 948 (Nov 2026)

OpenAI's agent pipeline found 948 unique Linux kernel privilege-escalation flaws and made 1,169 working root exploits in one run (reported Nov 5, 2026).

Jemiboo Magazine

· 5 min read

Read in other languages한국어Español简体中文繁體中文日本語

Key takeaways (as of November 6, 2026)

  • Help Net Security reported that the run cost $1,213,430 in API fees, about $1,279 per exploit, and used an OpenAI cybersecurity model that has not yet been released.
  • In its Android security bulletin of November 2, 2026, Google split the security patch level into 2026-11-01 and 2026-11-05 and grouped kernel component vulnerabilities under 2026-11-05.
A server drawer opened in a data center, with an unlocked padlock

An AI agent pipeline has found 948 vulnerabilities in the Linux kernel that let an attacker seize administrator (root) privileges, all in a single run. The research comes from OpenAI, and Help Net Security and SecurityWeek reported it in detail on November 5. In the same week, Google (November 2) and Samsung Electronics (November 3) published their November security updates.

How 948 Linux kernel vulnerabilities were found

The Linux kernel is the core of the operating system. Servers and cloud platforms run on it, and so do Android smartphones. Help Net Security cited this breadth of use as the reason OpenAI chose Linux. What the agents found are local privilege escalation (LPE) vulnerabilities: paths that let someone logged in with an ordinary user account climb to administrator privileges. The work was divided among agents assigned to separate stages. A reconnaissance agent first narrowed about 7% of the kernel code down as areas worth examining. An analysis agent produced 2,308 candidate bugs in 12 hours, and the next stage reproduced 1,724 of them (74.7%) over 30 hours. The test environment was the default desktop configuration of a widely used distribution, with kernel protections such as KASLR, SMEP and SMAP switched on.

StageResult (count)Time taken
Candidate bugs found2,30812 hours
Successfully reproduced1,724 (74.7%)30 hours
Information-leak exploits (PoC)291—
Working root exploits1,169—
Unique vulnerabilities (exploits grouped by flaw)948—

Figures are as reported by Help Net Security on November 5, 2026, and were compiled by OpenAI itself. The 948 figure counts exploits stemming from the same flaw as one; SecurityWeek described it as the number of separate patches written, one for each bug.

About $1,279 per exploit, and how that compares with contest prizes

The run cost $1,213,430 in API fees, which works out to roughly $1,279 per working exploit. That is far below what the security industry pays. The hacking contest Pwn2Own has offered $20,000 for a Linux kernel privilege escalation, and TyphoonPWN $70,000. On that basis, SecurityWeek noted, the bugs found were worth 16 to 55 times the cost of the scan. There was also a test modeled on kernelCTF, the kernel exploitation contest run by Google. Over 112 hours and at an API cost of $32,936, it found 21 privilege escalations.

ItemAmount (USD)Note
OpenAI single-run API cost1,213,4301,169 exploits
Cost per exploitabout 1,279OpenAI's count
Pwn2Own kernel privilege-escalation prize20,000Hacking contest
TyphoonPWN kernel privilege-escalation prize70,000Hacking contest
kernelCTF-style test API cost32,936112 hours · 21 privilege escalations

Amounts are as reported by Help Net Security on November 5, 2026; prizes are what the contests have offered to date. The limitations were disclosed as well. In an early run, the exploitation stage reported 200 successful root exploits, but 20 of them reached administrator privileges through other zero-days rather than the vulnerability they had been assigned. After reviewing a sample of 166 cases, OpenAI classified 10 of the exploitation techniques the agents used as novel.

A smartphone and laptop updating on a desk

How the vulnerabilities are being fixed

Along with the exploits, the agents drafted a patch and a severity assessment for each bug. Security researchers reviewed them one by one before handing them to kernel maintainers. At the kernel community's suggestion, the most dangerous flaws went first; the rest are being submitted in stages through Patch the Planet, OpenAI's open-source security program. Help Net Security described the status over the ten days the scan ran: 61 flaws had patches merged into the upstream kernel, 43 had patches in staging, and 299 were under public discussion. SecurityWeek, relaying OpenAI's account, summarized the same point as more than 100 fixed or staged through patches from other contributors and about 300 under public discussion. Not all 948 have been fixed. Both outlets reported the researchers' warning that on systems slow to patch, obtaining a single low-privilege account is, for a well-resourced attacker, effectively the same as having administrator privileges.

Date (2026)OrganizationDetails
November 2GoogleAndroid November security bulletin published (patch levels 2026-11-01 · 2026-11-05)
November 3Samsung ElectronicsGalaxy November security update bulletin published
November 5Samsung ElectronicsCVE list in the November bulletin updated
November 5Help Net Security · SecurityWeekReports on OpenAI's kernel research

What changes for users and businesses in Korea

The closest point of contact is the phone. In its November 2 bulletin, Google split the security patch level into two stages, 2026-11-01 and 2026-11-05, and placed kernel component vulnerabilities in the later one. Google wrote that the most severe of them could lead to remote privilege escalation. Samsung Electronics, the South Korean maker of Galaxy smartphones, published its own Galaxy bulletin on November 3, saying Google's patches are included in the November package. It listed 24 entries for Samsung's own vulnerabilities (SVE). Neither bulletin says whether any of the 948 flaws OpenAI found are in this month's release. Still, these monthly updates are how kernel patches reach users. For Korean companies and public institutions running Linux servers, it becomes a question of scheduling. The faster AI finds flaws, the more any delay in applying distribution kernel updates turns into a period of exposure.

Frequently asked questions

How many Linux kernel vulnerabilities did OpenAI find?

Reports on November 5, 2026 said OpenAI's agent pipeline found 948 unique privilege-escalation vulnerabilities and produced 1,169 working root exploits in a single run. The figures were compiled by OpenAI, and not all of the flaws have been fixed.

Are Android phones affected by Linux kernel vulnerabilities?

Android runs on the Linux kernel, so kernel vulnerabilities extend to phones. Google's November bulletin groups kernel flaws under the 2026-11-05 patch level, and you can check your phone's security patch level date in the software information section of its settings.

Has OpenAI released the model used in the research?

Help Net Security and SecurityWeek reported that the research used an OpenAI cybersecurity model that has not yet been released. Neither report gave the model's name or a date for general availability.

Readers can check in their phone's software information settings whether the security patch level reads 2026-11-05. For servers, the first item on the checklist is whether a distribution kernel update dated on or after that day has arrived.

Sources

  1. Help Net Security, OpenAI's upcoming AI model built 948 working Linux root exploits in a single run (Nov 5, 2026) — helpnetsecurity.com
  2. SecurityWeek, AI Agents Got Root on Linux via 948 Kernel Flaws in OpenAI Research (Nov 5, 2026) — securityweek.com
  3. Google, Android Security Bulletin, November 2026 (Nov 2, 2026) — source.android.com
  4. Samsung Mobile Security, November 2026 security update (Nov 3, 2026; updated Nov 5) — security.samsungmobile.com

OpenAIAI securityAI agentsLinux kernel vulnerabilitiesLinux kernelprivilege escalation

More from the same day

More in Tech