What happens when an AI finds 948 ways to root Linux?

One run, about $1.2 million, and a pile of bug reports the kernel's maintainers now have to read.

Think about what finding a single good privilege escalation bug in the Linux kernel normally takes. Weeks of a skilled researcher's time. A contest like Pwn2Own paying out around $20,000 for one working exploit, or up to $70,000 at another event. Now meet a number: 948.

On November 4, OpenAI said an autonomous pipeline running an upcoming, unreleased cyber model had validated and written patches for 948 unique local privilege escalation vulnerabilities in the Linux kernel. One run, in mid-October. About $1.2 million in API costs.

How it works, as a funnel

Think of it as a funnel with five stages. Scouting agents pick out areas of a codebase with more than 40 million lines, roughly five thousand of them. Analysts and triage steps flag candidates, about 2,300 in all. Of these, 1,724 were reproduced. Exploit agents then built working attacks, tested against defences like KASLR and SMEP, and drafted candidate patches. After grouping duplicates, you get 948 unique ones. There were also 291 information-leak proofs of concept.

A human researcher reviews exploits, patches and severity before anything goes to the kernel's maintainers. And in a sample of 166 exploits, OpenAI and outside kernel experts agreed that ten techniques were new.

Ok, but if this is so cheap and so good, isn't that exactly what the bad guys want?

Yes, and that is the whole point

OpenAI compares the cost to old bounty payouts: apply the historic prices and the 948 exploits are worth between $18.9 and $66.3 million, sixteen to fifty-five times what the run cost. That is a defender's argument. It is also an attacker's spreadsheet. The same pipeline run by someone else would not come with a coordinated disclosure plan. OpenAI says it is working with kernel maintainers and security partners to get these closed before anyone copies the playbook.

Which brings us to the bottleneck nobody talks about: the people on the receiving end. Finding bugs just got dramatically cheaper. Fixing them, reviewing patches and shipping them to billions of devices did not.

Read the fine print

Every number here comes from OpenAI. There is no public list of CVEs or commits yet to check them against. The report itself even contains two exploit counts: 1,169 working escalation exploits in one place, 948 unique ones in the summary. And a small share, 3.1 percent, reflected incorrect analysis by the model.

So treat it as a serious claim, not a settled fact. If it holds, the interesting race is not between humans and AI. It is between AI that finds bugs and AI that fixes them, with a very small group of maintainers in the middle.