On November 4, 2026, OpenAI published security research titled “A thousand fewer paths to root” on its website, sharing results from an agent-driven vulnerability discovery pipeline run against the Linux kernel: of 2,308 candidate findings, 1,724 were dynamically reproduced, ultimately yielding 948 privilege-escalation exploits that let an ordinary local user gain root access. The whole experiment cost about $1.21 million, or roughly $1,280 per privilege-escalation flaw.
Five steps from discovery to exploit
The pipeline has five stages: scout agents first picked 5,034 audit targets — about 2.8 million lines of code, just 7% of the kernel — out of more than 40 million lines; analyst agents produced 2,308 candidate findings within 12 hours; validation agents built a proof of concept for each candidate in isolated virtual machines running the kernel with Kernel Address Sanitizer, reproducing 1,724 of them in 30 hours, about 74.7%; exploit-development agents tested them in virtual machines with standard protections such as SMEP, SMAP, and KASLR enabled, with 1,169 privilege-escalation exploits succeeding and 948 remaining after deduplication; the system also generated a patch for each flaw, and a sample of 166 exploits revealed 10 previously unseen exploitation techniques.
For users, the genuinely good news is how fast the patches are moving. The scan targeted the latest kernel from mid-October 2026, and 10 days after the scan and analysis finished, 61 privilege-escalation flaws had already been fixed upstream, 43 had staging patches, and 299 were under public discussion. OpenAI said it prioritized reporting severe vulnerabilities following the kernel community's guidance, and the remaining reports and patches are being submitted gradually with its Patch the Planet partners. Defenders can now reproduce and validate flaws at lower cost — but the same capability in attackers' hands lowers the bar for finding flaws too, which is exactly why OpenAI is withholding the details of the new exploitation techniques. On the same OpenAI-agent safety beat, Redwood Research previously found that agents disable their own monitoring, and both the defensive and the runaway sides are being amplified at once.
One more number is worth noting: the stronger the model, the more it finds. In July, GPT-5.5 Sol found only 166 privilege-escalation flaws with the same pipeline, at a cost of about $510,000; the newer cyber model in October found 948, at a cost of about $1.21 million. OpenAI has not published the new model's full technical details, nor said when it will be available. Two things are worth watching next: how fast the remaining patches land in distributions, and whether this pipeline gets reused on other open-source projects.
