On Nov. 4, 2026, OpenAI published “A thousand fewer paths to root,” describing an autonomous vulnerability discovery, exploitation, and patching pipeline powered by an upcoming cyber model. OpenAI frames the work as research under Patch the Planet and Daybreak, not a ChatGPT or Codex product launch. For related coverage, see Google Unveils Gemini 4 Carbon for Cyber Defenders First and CrowdStrike Launches AI Kill Switch to Stop Rogue Agents.
Key facts
| Item | Detail (per OpenAI) |
|---|---|
| Published | Nov. 4, 2026 |
| Scope | Research pipeline; upcoming cyber model (not described as publicly released) |
| Mid-October 2026 run | 1,724 vulnerabilities dynamically reproduced; PoCs for 291 information-leak issues and 948 working local privilege escalation (LPE) exploits on a commonly used Linux distribution |
| Run cost | $1,213,430 in API credits (~$1,279 per LPE exploit) |
| Target environment | Stock desktop OS on VMs (not weakened); threat model aligned with Pwn2Own / TyphoonPWN |
| kernelCTF-adapted run | 25 info-leak PoCs and 21 LPE exploits; ~$32,936 API credits; if accepted, ~$1.5M bounty (~45× cost) |
| Human gate | Researchers review before reporting; hundreds of critical reports and patches already sent |
Pipeline stages
OpenAI describes five linked stages: discovery, dynamic validation, exploitability assessment, patch generation, and human review. Scout agents surfaced 5,034 audit targets covering roughly 2.8 million lines of code, about 7% of the kernel. Analyst agents produced 2,308 candidate findings after deduplication.
Dynamic validation reproduced 1,724 issues, about 74.7% of candidates; OpenAI says only 3.1% of analyses were incorrect. The 948 LPE exploits and 291 information-leak PoCs are pipeline outputs OpenAI is reporting alongside patches—not a claim that every exploit is a confirmed upstream zero-day.
Mid-October Linux desktop run
In the mid-October 2026 campaign on a commonly used Linux distribution, OpenAI says the pipeline dynamically reproduced 1,724 vulnerabilities and built proof-of-concept exploits for 291 information-leak bugs and 948 working LPE chains. Target virtual machines ran a stock desktop operating system without deliberate weakening, under a threat model OpenAI compares to Pwn2Own and TyphoonPWN.
OpenAI reports total API spend of $1,213,430 for that run, or about $1,279 per LPE exploit. OpenAI also presents a hypothetical valuation: applying historical Pwn2Own ($20,000) and TyphoonPWN ($70,000) LPE bounty levels to the 948 exploits would imply $18.9 million to $66.3 million—roughly 16× to 55× the pipeline cost. OpenAI presents that as a historical-bounty thought experiment, not revenue OpenAI collected.
OpenAI says agents used at least 10 novel Linux exploitation techniques in a sample of 166 exploits, phrased as “to the best of our knowledge” in the source post; here that means OpenAI believes the techniques are novel to its knowledge.
kernelCTF campaign
OpenAI adapted the pipeline for kernelCTF and reports 25 information-leak PoCs and 21 LPE exploits from a campaign costing about $32,936 in API credits. OpenAI says acceptance would imply roughly $1.5 million in bounty value, about 45× the run cost. For context, OpenAI notes 26 confirmed zero-day LPEs accepted by kernelCTF in 2026 at the time of writing.
July versus mid-October models
OpenAI contrasts a late July 2026 run with GPT-5.6 Sol that produced 166 unique LPEs against the mid-October run with the newer cyber model yielding 948 unique LPEs—more than five times as many novel LPEs despite roughly three months of hardening between the scans.
When the autonomous scan and manual analysis finished after 10 days, OpenAI reports 61 LPEs had already been patched upstream, 43 had staging patches, and 299 were being reported and discussed publicly.
Reporting and partners
OpenAI states human researchers review findings before reporting. OpenAI says hundreds of critical vulnerability reports and patches have already been sent and that it is working with Patch the Planet partners. The post situates the pipeline inside Daybreak and Patch the Planet defensive research, separate from consumer model releases.
What it means
Analysis (ours). OpenAI is publishing scale metrics for an agentic offensive-defensive loop—discovery through patches—rather than shipping the cyber model as a general product. The dollar-per-exploit and hypothetical bounty multiples are OpenAI’s way to compare API spend against human contest economics, with explicit caveats that bounties are illustrative. The gap between 948 pipeline LPEs and kernelCTF’s accepted zero-day count underscores that validation, disclosure, and maintainer acceptance still gate what counts as fixed in the wild.
What to watch
Whether OpenAI releases the upcoming cyber model beyond research use, how many mid-October LPE reports land as accepted upstream fixes, kernelCTF outcomes for the 21 LPE submissions, and whether other vendors publish comparable autonomous patch pipelines with public cost data.