OpenAI says AI built 948 Linux root exploits in one run

On November 4, OpenAI said a prototype autonomous security pipeline built working root-access exploits for 948 distinct Linux kernel vulnerabilities in a single run. The run used an unreleased cyber model and cost $1,213,430 in API credits, which works out to about $1,279 per exploit, according to OpenAI's research post. The company presents it as defensive work. It also says it shows that finding bugs is no longer the hard part. Checking, ranking and patching them is.openai.com

From 2,308 leads to 948 exploits

The scan ran in mid-October 2026 against the latest Linux kernel master branch. Scout agents narrowed the kernel's more than 40 million lines to 5,034 audit targets, about 2.8 million lines or roughly 7% of the codebase, Unite.AI reported. Analyst agents produced 2,308 candidate findings in 12 hours. Validation agents then reproduced 1,724 of them inside isolated virtual machines within 30 hours, Interesting Engineering reported.unite.ai+1

The exploit stage produced 291 information-leak proofs of concept and 1,169 working local privilege escalation exploits. These let an unprivileged user gain root. Variant analysis traced those exploits back to 948 distinct vulnerabilities, each with its own candidate patch. The test machines ran a stock desktop system with standard protections such as SMEP, SMAP and KASLR switched on, OpenAI said.openai.com

In a sample of 166 exploits, researchers found at least 10 exploitation techniques that they believe had not been documented before, and outside kernel experts agreed. OpenAI did not publish full details "for safety reasons."

Patches and human review

Human researchers review each exploit, patch and proposed severity before anything goes to Linux maintainers. OpenAI says it has sent hundreds of reports and patches for critical flaws and is working with partners in its Patch the Planet initiative to handle the rest, Interesting Engineering reported. The 948 fixes are candidates, not merged patches. By the end of the 10-day scan and analysis, 61 of the vulnerabilities had been patched upstream, 43 had staging patches, and 299 were already being discussed in public, Unite.AI noted.interestingengineering.com+1

Review also caught a problem. In an early run, 20 of 200 reported successful exploits got root through a different zero-day than the one the agent was assigned. OpenAI responded by adding an independent check of each exploit's path.

A fivefold jump in three months

In late July, GPT-5.6 Sol found 166 unique privilege escalation flaws for $513,690. The October model found more than five times as many, even after three months of kernel hardening in between, OpenAI said. Priced at past Pwn2Own and TyphoonPWN bounty rates, the 948 exploits would be worth a hypothetical $18.9 million to $66.3 million.openai.com

The company's warning to defenders is blunt: "having access to a low-privileged account on an operating system, that isn't up-to-date on its latest available patches should be considered functionally equivalent to root access for a well-resourced threat actor."