948 patches, with the cost stated

The OpenAI news feed places A thousand fewer paths to root at Wed, 04 Nov 2026 08:00:00 GMT, which is 16:00 in Beijing on Nov. 4, the inclusive start of this edition's window. The article page says only November 4, 2026. The summary says one run with an upcoming cyber model validated and generated patches for 948 local privilege-escalation vulnerabilities in Linux, at an API cost on the order of $1.2 million. A later section is more exact about that mid-October 2026 run: 1,724 vulnerabilities were dynamically reproduced, proofs of concept were generated for 291 information leaks, and 948 working local privilege-escalation exploits were produced. The total was $1,213,430 in API credits, or $1,279 per local privilege-escalation exploit. The exploit-development section then reports 1,169 working local privilege-escalation exploits, and variant analysis produced 948 distinct patches. The post uses 948 as the count of unique local privilege escalations. The figures 1,169 and 948 sit on the same page. They are not the same count.

The target virtual machines ran a stock desktop operating system, not a weakened target, under the threat model used by competitions such as Pwn2Own and TyphoonPWN. The post applies the historical bounty amounts those programs offered for a Linux kernel local privilege escalation, $20,000 and $70,000, to the 948 exploits. That arithmetic produces a hypothetical valuation of $18.9 million to $66.3 million, about 16 to 55 times the pipeline's cost. The post labels the valuation hypothetical. It also treats the reported costs as an upper bound, not a figure that has already been optimized. The text says the work measured method and model capability rather than cost, and that enough harness engineering could lower the cost of each stage. A person reviews the exploit, the patch, and the proposed severity before a report goes out. Hundreds of reports and patches for critical issues have already been sent. The rest are to go out gradually with Patch the Planet partners.

On an adapted kernelCTF setup, SLAB_VIRTUAL was dropped from the configuration because it was incompatible with that program's mitigation. The run produced 25 information-leak proofs of concept and 21 local privilege-escalation exploits. The comparison figure is the 26 confirmed zero-day local privilege escalations kernelCTF had accepted in 2026 at the time of writing. The campaign lasted 112 hours and cost roughly $32,936 in API credits, with vulnerability discovery accounting for 83.7 percent of that API cost. If the 21 findings were accepted by the program, the award would be around $1.5 million, about 45 times the cost. Acceptance has not happened. The multiple is conditional. In late July 2026, a GPT-5.6 Sol run found and exploited 166 unique local privilege escalations at an end-to-end API cost of $513,690. The post says the newer model in the October run found more than five times as many novel local privilege escalations as that July run. Both scans used the latest Linux kernel master branch at the time.

When the autonomous scan and the manual analysis finished after 10 days, 61 local privilege escalations had already been patched upstream, 43 had staging patches, and 299 were being reported and discussed in public. That trio is more precise than the summary's "over 100" patched or staging and "about 300" under discussion, so the body uses the trio. During validation, only 3.1 percent of the remaining findings were classed as incorrect analysis, meaning the report assumed a condition that could not occur. The post attributes most of the rest that failed to reproduce to a need for higher privileges, to the time budget, or to trigger conditions that were hard to hit. An early exploitability run marked 200 successes. Human review found that 20 of them had reached the highest privilege through a different zero-day than the one assigned. The task constraints were then tightened, and an independent check was added so that the final result had to use the assigned vulnerability. The same kernelCTF run also hit a vulnerability in QEMU TCG. The post says it had already been fixed in upstream QEMU, but not in the release build used for the experiment.

The post says a sample of 166 exploits contained at least 10 techniques that were new to the authors and to outside kernel experts, and that the technical details are withheld for safety. This column does not relay those techniques, and it does not break discovery, reproduction, exploitation, or patch writing into steps that could be followed. A separate experiment in the post, which supplied a catalog of existing techniques and then compared success rates, is also left unreproduced. What is public, and what the post itself qualifies, are the counts, the costs, the hypothetical valuation, the bounty multiple that still depends on acceptance, and the review that found 20 results using a different vulnerability. The cover is the post's own social card. It shows the title, not this table of counts.