Skip to content
Hacker Posts

OpenAI says an unreleased cyber model generated 948 Linux LPE exploits for $1.2M

OpenAI's 'A thousand fewer paths to root' claims an upcoming cyber model reproduced 1,724 Linux kernel bugs and produced working exploits for 948 distinct privilege escalations at $1.2M in API cost.

Published 3 min read
researchai-securitylinuxopenai

OpenAI has published figures for a pipeline that used an unreleased "cyber model" to find, reproduce and write exploits for Linux kernel privilege-escalation bugs at scale. The primary is OpenAI's own writeup, A thousand fewer paths to root (November 4), accompanied by a preprint. This is a vendor capability demonstration with self-reported numbers — read it as such.

What they claim

In a mid-October run, the pipeline reportedly:

  • analyzed 2,308 candidate Linux kernel findings,
  • dynamically reproduced 1,724 vulnerabilities,
  • produced 1,169 working exploits, which variant analysis collapsed to 948 distinct local privilege escalations, plus 291 information-leak proofs of concept.

Total API cost was $1,213,430 — roughly $1,279 per LPE. OpenAI contrasts this with a July run on an earlier model it calls GPT-5.6 Sol, which produced 166 unique LPEs for $513,690. The current model is unnamed, described only as "upcoming."

Why a practitioner should care

The interesting number isn't 948 — it's the cost curve. Reproducing a kernel bug and producing a working LPE is exactly the labor that has gated exploitation: reading the patch, triggering the condition, and stabilizing a primitive into root. If that work drops toward $1,000 per bug and runs unattended, the gap between "a CVE exists" and "a reliable exploit exists" narrows for anyone who can pay the API bill. The defensive read is that your patch SLA for kernel LPEs is now competing against an automation that scales linearly with budget.

Caveats — read these before the headline

  • Self-reported, pre-release. The model isn't available, the figures aren't independently reproduced, and OpenAI has an incentive to make the demo land. "948 exploits" is OpenAI's count against candidate findings it selected.
  • N-days, not 0-days. These target known kernel findings — reproduction and exploitation of existing bugs, not discovery of novel ones. That is still consequential, but it is not "AI found 948 zero-days."
  • LPE, not remote. Local privilege escalation requires a foothold first. This lowers the cost of post-exploitation, not initial access.
  • Exploits for patched bugs are dual-use. OpenAI says it is coordinating with Linux maintainers and "Patch the Planet" partners on reporting and fixes; the same corpus is a roadmap for an attacker who obtains it.

What to do today

  1. Treat kernel-LPE patch latency as a measurable risk, not a backlog item. Prioritize local-priv-esc fixes on multi-tenant and internet-adjacent hosts.
  2. Invest in the controls that survive a reliable LPE: workload isolation, least-privilege service accounts, and runtime detection of privilege transitions — not just patching.
  3. Don't restructure your program around a benchmark you can't reproduce. Track whether independent researchers replicate the throughput before treating it as a baseline.

Context

This is the same thread as Trail of Bits backporting a kernel CVE with Codex: AI is moving from bug-finding assistance toward end-to-end exploit production, and the defensive question shifts from "can a model find it" to "how fast can it weaponize what's already public." The honest answer this week is a vendor's own number — worth watching, not yet worth believing wholesale.

Related stories