A thousand new paths to root, in thirty hours

November 5, 2026 · SPR{K3 Research

On November 4, OpenAI published a results post titled "A Thousand Fewer Paths to Root", describing an internal agentic pipeline that spent thirty hours of wall-clock time finding and exploiting Linux kernel vulnerabilities. The output is specific. The pipeline looked at 2,308 candidate findings, dynamically reproduced 1,724 of them, and built 948 working local-privilege-escalation exploits against a commonly deployed Linux distribution — plus 291 proof-of-concept exploits for information-leak bugs. Total cost in API credits: $1,213,430 (The Hacker News, Nov 5).

That is roughly $1,279 per working root exploit, measured at the frontier lab's own internal rate.

The headline is "fewer paths," because OpenAI reported the findings to kernel maintainers under its Patch the Planet initiative and submitted hundreds of patches. Read it the other way and it is also a thousand new ways in, cataloged and reproducible, all against the same OS kernel most of the world runs.

What the numbers describe

Per the post and the Interesting Engineering writeup:

The novel-technique count is the part worth sitting with. The pipeline is not just rediscovering primitives from training data. Across four categories, it produced ten patterns the public record did not have before.

The numbers around the numbers

OpenAI's post lands next to industry data that points in the same direction.

And last week, on November 3, the Xen Project disclosed CVE-2026-102703 and CVE-2026-102704 — two grant-copy flaws that let a malicious guest VM read and write other VMs' memory. The Xen Project credits Anthropic with finding them using Claude. Ada Logics validated the report; the CodeAnt AI Security Research Team independently reported the same flaws. Two frontier labs, two different agentic pipelines, two different critical-infrastructure codebases, same week.

What OpenAI is now saying out loud

From the writeup, OpenAI's own framing for defenders: "A low-privileged account on an unpatched OS should be treated as functionally equivalent to root for a well-resourced attacker."

That is a frontier lab, in public, assigning the CVE-backlog defender posture a specific operational meaning. If you have a user account and you are not patched to the day, assume the attacker is already root. Not eventually. Now.

The lab is also saying, in the same post, that patch generation has to keep pace with discovery, because the same pipeline that writes the patch can write the payload. That symmetry is not new in principle — the research community has been writing about offensive-defensive AI parity since 2024 — but the public, per-run, per-cost number is new.

What this means for defense

Three things change when the per-exploit cost is public.

  1. The attacker arithmetic is now legible. A $1.2M pipeline run produced 948 working LPEs. Adversary-side replication drops that cost further — infrastructure is cheaper, review step gets skipped, patches get kept rather than reported. The GTIG trend-line (disclosures up 113% in eight months) is what that economic reality looks like in the vuln feed.
  2. The signature-and-CVE treadmill was already losing. Now it is losing faster. When ten novel exploitation techniques land in one run, the gap between "patch released" and "patch deployed" is where the attacker lives. Supply Chain Worms 2026 coverage at Dark Reading has been pointing at the same gap all year.
  3. The defender posture OpenAI just published is the right one. Assume compromise at the user-account layer. Watch what happens after. The sequence of actions after "low-priv user got a shell" is the only layer where the defender has leverage the attacker did not pre-generate.

The defense against a thousand new paths to root cannot be a thousand new patches on the host. It has to be observation of the attempt — the syscall sequence, the escalation pattern, the lateral move — that an AI-generated exploit still has to run in order to become real. Exploitation is a program; a program is a sequence; a sequence is observable.

The takeaway

Industrial-scale exploit generation at tier-1-lab quality is now a thing people publish numbers about. A pipeline that spends thirty hours and $1.2M to produce 948 working root exploits is a demonstration that per-bug patching, as the primary defense, is a game whose end state is already here.

The defense that works is the one that watches the sequence that follows. The defense that doesn't work is the one that races the patch cadence.

OpenAI itself just said as much, in different words.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.