OpenAI says an upcoming cyber model found and patched 948 Linux root bugs in one $1.2 million run

OpenAI says a pipeline built on an upcoming, unnamed cyber model found, exploited and wrote patches for 948 distinct local privilege escalation bugs in the Linux kernel in a single run, for $1,213,430 in API credits. That is about $1,279 per bug that lets an unprivileged local user take root. OpenAI's security team published the results on November 4.

WHAT OPENAI SAYS IT DID

The run, in mid-October against the latest kernel, went in stages. Scout agents picked 5,034 audit targets, about 2.8 million lines or 7 percent of the kernel. Analyst agents produced 2,308 candidate findings, and triage agents reproduced 1,724 of them, about three quarters, in isolated machines. Exploit agents then got root on 1,169 of them on a stock desktop distribution with its standard protections on (SMEP, SMAP, KASLR, hardened usercopy), and 948 distinct patches came out of those. OpenAI says only 3.1 percent of the candidate findings were wrong analysis, and that humans reviewed everything before reporting.

THE COMPARISON

OpenAI's own numbers: a GPT-5.6 Sol run in late July found and exploited 166 unique bugs of the same kind for $513,690, so the newer model found more than five times as many with three months of kernel hardening in between. On a version of Google's kernelCTF setup, which has less attack surface, the pipeline produced 21 root exploits in 112 hours for about $32,936; kernelCTF has accepted 26 confirmed zero-day root exploits all year. OpenAI notes that it dropped one kernelCTF mitigation (SLAB_VIRTUAL) because the latest kernel did not support it.

WHAT IS SOURCED

OpenAI's post, read directly. It says hundreds of reports and patches for critical bugs have already gone to the kernel maintainers, starting with the most serious as the kernel community asked, and the rest are going out with its Patch the Planet partners. Within the 10 days the scan and review took, 61 of the bugs had been patched upstream by others, 43 had staging patches and 299 were being discussed publicly. OpenAI says its agents used at least 10 exploitation techniques that outside kernel experts had not seen, and it is withholding the details. It also says that in an early run, 20 of 200 reported successes reached root through a different zero-day than the one assigned, and it has since added a check for that.

WHAT WE DO NOT KNOW

Which model it is. OpenAI calls it only "an upcoming cyber model". On October 23 it referred to "our upcoming GPT-6 Astra Cyber model", but this post does not say they are the same, and @LuminaBench's claim on X that it is the cyber model that leaked earlier is unconfirmed. There is no date, price or access policy. The Information reported on November 4 that new cyber-focused OpenAI models could come as soon as Friday, November 6; OpenAI has not confirmed that. Every number here is OpenAI's own and has not been checked independently. That includes its estimate that the 948 bugs would have been worth $18.9 million to $66.3 million at past Pwn2Own and TyphoonPWN bounty rates.