OpenAI’s upcoming cyber model found 948 Linux flaws that can grant root access
OpenAI says an automated pipeline using an upcoming cyber model found 948 distinct Linux kernel flaws that could let a local user gain root access, and generated candidate patches for human review. The mid-October run cost €1.1 million ($1,213,430); OpenAI says it has sent hundreds of reports and patches to maintainers.
Bottom line — OpenAI reports 948 distinct flaws, but the model remains unreleased and the results have not been independently verified.
Go deeper 6
-
OpenAI says the system produced 2,308 candidate findings, reproduced 1,724 and generated 1,169 working exploits before variant analysis narrowed them to 948 distinct flaws.
-
The pipeline examined 5,034 selected targets—about 2.8 million lines, or 7% of the kernel—rather than scanning the entire codebase, according to OpenAI.
-
OpenAI says the October run found more than five times as many unique flaws as a July run, which found 166 for €473,000 ($513,690).
-
The company says 61 flaws had been patched upstream and 43 had staging patches after 10 days; 299 more were being publicly discussed.
-
OpenAI reports that a sample of 166 exploits used at least 10 techniques that third-party kernel experts also considered new; it withheld their details.
-
An earlier test produced 200 apparent successes, but OpenAI says human review found 20 had gained root through a different flaw than the one assigned.