Back

OpenAI Reports 948 Linux Privilege-Escalation Exploits

lite News Desk 2 min read

OpenAI says a prototype system for investigating Linux kernel vulnerabilities generated 948 distinct local privilege-escalation exploits, along with candidate patches, during a mid-October 2026 run. Using an upcoming cyber model, the system examined 2,308 candidate findings and reproduced 1,724 vulnerabilities through dynamic testing, according to the company.

The work concerns flaws that let someone with limited access to a machine gain root access, which OpenAI describes as the highest level of privilege on Linux. The company published the results on November 4, 2026, in its research disclosure, A thousand fewer paths to root.

OpenAI said 61 local privilege-escalation vulnerabilities had received upstream patches by the end of its 10-day scanning and manual-analysis effort. Another 43 had patches in staging, while 299 were under public discussion. The disclosure does not establish that those status figures account for all 948 distinct vulnerabilities from the run.

The company said it had submitted hundreds of reports about vulnerabilities and patches addressing critical issues, following guidance from the Linux kernel community. Work on additional reports and patches was continuing with partners in Patch the Planet, according to OpenAI.

For discovery, the prototype selected portions of kernel code that users with limited privileges could access, then searched for potential flaws. OpenAI said its analysis concentrated on candidates that needed no special capabilities and showed measurable evidence of memory corruption. Reproduction testing used Kernel Address Sanitizer inside isolated virtual machines.

Once flaws had been validated, the system assessed whether they could expose information or enable privilege escalation, according to the company. Exploit testing took place in virtual machines running a standard Linux distribution. OpenAI said SMEP, SMAP, KASLR and HARDENED_USERCOPY remained enabled during those tests; the tests did not depend on disabling those protections. The workflow also generated proposed fixes for review by human security researchers.

OpenAI put the run's API-credit cost at $1,213,430, averaging $1,279 for each distinct local privilege-escalation exploit. It said that estimate was specific to the models, target configurations and workflow used. Further engineering of the testing harness may lower the cost, according to the company.

The research extends an effort from June 2026 in which GPT-5.5 Cyber identified Linux kernel vulnerabilities and generated 24 local privilege-escalation exploits, OpenAI said. The later pipeline expanded that work into a process covering flaw discovery, validation through execution, assessment of exploitability and the generation of candidate patches.

* This article is based on publicly available information at the time of writing.

More from Artificial Intelligence

  1. OpenAI says Code Storage’s Git-based service will join Codex only if the acquisition closes, subject to regulatory approvals.

  2. Palantir said U.S. commercial revenue reached $984 million in the third quarter, exceeding U.S. government revenue for the first time.

  3. ADNOC’s planned field-development system will draw on decades of data from thousands of wells, with a target of producing plans within hours instead of months.

  4. The draft would cover authorised telecom entities that use or plan to use AI for services or network operations, according to the regulator.

  5. OpenAI plans a voluntary implementation in the coming months; the draft lets site operators set training preferences for specific paths.