Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

948 Linux exploits: what the pipeline actually did

Written November 6, 2026

OpenAI's pipeline reproduced 1,724 Linux vulnerabilities and built 948 working exploits for $1.21M. What the numbers prove — and what they don't.

The headline, unpacked

OpenAI's research post describes a prototype pipeline run against the latest Linux kernel development code in mid-October 2026: 2,308 candidate findings, 1,724 reproduced vulnerabilities, and 948 distinct local privilege escalation exploits: working code that lets an unprivileged local user gain root access, in a controlled research environment. The campaign cost $1,213,430 in API credits. Every number in this paragraph is OpenAI's own; the pipeline has not been independently reproduced.

The funnel matters more than the final number. Each stage filtered the last, and each stage cost money. A candidate finding is not a vulnerability, a reproduced vulnerability is not an exploit, and an exploit is not an accepted patch. The post is careful about those distinctions, and the headline usually is not.

The five stages

The pipeline has five stages, with human review at the end:

  1. Vulnerability discovery. Scout agents read the kernel and picked 5,034 audit targets — about 2.8 million lines of code, roughly 7% of the kernel — based on reachability by low-privileged users and security relevance. Analyst agents then scanned those targets and produced 2,308 candidate findings after deduplication, in 12 hours. Scouting cost $3,426 in API credits; analysis cost $316,549.
  2. Validation. A triage agent generated a proof of concept for each candidate and ran it in an isolated virtual machine running a kernel with Kernel Address Sanitizer, refining the proof until it reproduced the behaviour or ran out of ways to trigger it. Within 30 hours, 1,724 vulnerabilities reproduced, about 74.7% of the candidates. The rest mostly needed elevated privileges, failed to reproduce inside the time budget, or, in 3.1% of cases, reflected incorrect analysis by the model. Validation cost $73,330.
  3. Exploitability assessment. A screening agent narrowed the 1,724 validated vulnerabilities to 313 information-leak candidates and 1,238 local privilege escalation candidates. Exploit-development agents then tested them in virtual machines with the protections of a standard Linux distribution enabled (SMEP, SMAP, KASLR, and HARDENED_USERCOPY) and produced 291 information-leak proofs of concept and 1,169 working local privilege escalation exploits. Variant analysis collapsed those into 948 distinct exploits, with 948 distinct patches developed for the vulnerabilities behind them. Screening cost $44,194; exploitation cost $775,929.
  4. Patch generation. For each validated vulnerability the system generated a candidate patch, following the coding conventions of nearby code and checking for variants of the same flaw. Early patches were functionally correct but often bloated, non-idiomatic, or incomplete; additional instructions and iterations improved them.
  5. Human review. Security researchers reviewed the exploit, patch, and proposed severity before anything was reported to maintainers. The pipeline bundles a root cause analysis, a minimal crashing proof of concept, a candidate patch, and a full exploit for that review.

What it cost

The total was $1,213,430 in API credits, which works out to about $1,280 per working exploit across the 948. That average hides a skewed distribution: a successful exploit cost an average of $359 in credits to develop, while the 69 unsuccessful exploitation attempts cost $3,922 each on average — a few hard-to-exploit vulnerabilities, such as race conditions, burned far more than the successes. OpenAI notes its costs are an upper bound, and that more work on the pipeline could reduce them.

For scale, the post applies historical bug-bounty amounts to the 948 exploits as a hypothetical: at $20,000 per Linux kernel local privilege escalation exploit (Pwn2Own's historical figure) or $70,000 (TyphoonPWN's), the haul would be worth $18.9M to $66.3M — roughly 16 to 55 times the pipeline's cost. That is OpenAI's hypothetical valuation, not money anyone received.

The comparison that shows the jump

The same test setup ran an earlier scan in late July 2026, when GPT-5.6 Sol found and exploited 166 unique local privilege escalation vulnerabilities for $513,690 in API credits. The mid-October run, powered by the newer cyber model, the model OpenAI has since named GPT-6 Cyber, found and exploited 948 for $1,213,430. Same setup, same target in the same state of development, 5.7 times the exploits at 2.4 times the cost. OpenAI's own takeaway is to rescan previously audited code as models improve: a stronger model finds what the earlier one missed.

The pipeline was also run against Google's kernelCTF program, a hardened kernel with a reduced attack surface: 25 information-leak proofs of concept and 21 local privilege escalation exploits in 112 hours, for about $32,936. For comparison, 26 confirmed zero-day local privilege escalation exploits had been accepted by kernelCTF in 2026 at the time of writing.

What the numbers do not prove

948 working exploits in a test environment is not 948 accepted patches. When the autonomous scan and manual analysis finished after 10 days, 61 of the local privilege escalation vulnerabilities had already been patched upstream, 43 had staging patches, and 299 were being reported and discussed publicly. OpenAI says it disclosed 304 critical findings to maintainers with candidate patches and is working through the rest with its Patch the Planet partners. Upstream acceptance is incomplete, and the kernel moves fast enough that findings go stale: the post's own takeaway is that automated triage has to revalidate against the latest code, including code merged after the original scan.

The pipeline also needed its own validation. In an early run, the exploitation stage reported 200 successful exploits, but human review found 20 of them had achieved root through a different zero-day than the one assigned. OpenAI tightened the task constraints and added an independent check confirming each final exploit used its assigned vulnerability. The lesson it draws is blunt: for autonomous security research, verify the path to a result as well as the result itself.

And the model did more than pattern-match. In a sample of 166 exploits, OpenAI says the agents used 10 new exploitation techniques — two new persistent elastic objects, two new page-grooming techniques, five new data-only techniques, and one arbitrary type confusion — confirmed as new by third-party kernel exploitation experts. The technical details were withheld for safety.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: GPT-6 Cyber: OpenAI's restricted security model · A malicious MCP server probed cloud credentials · Fingerprinting AI coding agents from 33,580 PRs