Roughly $1,279 per root exploit is the new number defenders have to price in.
OpenAI says an agentic pipeline running an unreleased cyber-focused model spent 30 hours in mid-October working a Linux kernel. It analysed 2,308 candidate findings, dynamically reproduced 1,724, and produced 948 working local privilege escalation exploits plus proofs of concept for 291 information leak bugs. Total API cost: $1,213,430.
The pipeline runs five stages: discovery of audit targets reachable by low-privileged users, sandboxed dynamic validation that keeps runtime evidence, exploitability assessment, patch generation, and human review before anything reaches maintainers.
Ten techniques nobody had published
To the best of the company’s knowledge, the agents invented at least ten novel Linux exploitation methods. They include two new persistent elastic objects for spraying the kernel cache, two page grooming techniques that improve reliability, five data-only paths to escalation, and an arbitrary type confusion.
The important half of this story is not the exploit count. Every one of those 948 findings came with a generated patch and a severity recommendation, which is the workload that historically outran kernel maintainers.
The defender’s problem
Google Threat Intelligence Group counts monthly disclosures more than doubling between January and August 2026, from 5,045 to 10,740, with high-risk findings up 167%. The UK’s NCSC said Tuesday that cheap AI tooling has almost certainly driven much of that.
Which leaves security teams facing a volume curve that no human triage process was sized for. The practical response is not to read faster. Shrink the surface that matters: cut unpatched kernel exposure, enforce a patch cadence that does not wait for a monthly window, and instrument privilege boundaries so an escalation attempt fails loudly.
OpenAI secures its own code faster and hands everyone else a higher flood. Both are true at once.
