Summary
On November 4, 2026, OpenAI published "A thousand fewer paths to root," a report on an automated pipeline that searched the Linux kernel for vulnerabilities, reproduced them, wrote exploits and drafted patches. In a mid-October run with an unreleased cyber model, the pipeline dynamically reproduced 1,724 vulnerabilities and produced 948 working local privilege escalation (LPE) exploits, each of which could take an unprivileged local user to root, for $1,213,430 in API credits, or $1,279 per exploit. The next day OpenAI named the model GPT-6 Cyber, classified it as Critical in cybersecurity under its Preparedness Framework, and released it to a select group of approved customers in its Daybreak Red program.
What Happened
The pipeline extended work OpenAI first described in June under its Patch the Planet initiative, when GPT-5.5 Cyber had developed 24 Linux LPE exploits to validate and report kernel bugs. OpenAI's November 4 post describes six stages: agents identify security-relevant parts of the kernel, statically scan them for candidate findings, reproduce the findings in an isolated environment, build a proof-of-concept exploit, generate a patch, and hand the result to a human security researcher, who reviews it before anything is reported to maintainers.
In the mid-October run against the latest kernel, the pipeline analyzed 2,308 candidate findings, reproduced 1,724 of them, and produced proof-of-concept code for 291 information-leak bugs and 948 working LPE exploits. The target virtual machines ran a stock desktop build of what OpenAI called "a commonly used Linux distribution," under the threat model used by the Pwn2Own and TyphoonPWN competitions. Applying those competitions' historical Linux LPE bounties of $20,000 and $70,000 to the 948 exploits, OpenAI put their hypothetical value at $18.9 million to $66.3 million, about 16 to 55 times the run's cost. On an adapted kernelCTF environment, a kernel configuration with reduced attack surface, the pipeline produced 21 LPE exploits for about $32,936; OpenAI noted that kernelCTF had accepted 26 confirmed zero-day LPE exploits so far in 2026.
OpenAI compared the run with an earlier scan of the same codebase. In late July, GPT-5.6 Sol found and exploited 166 unique LPE vulnerabilities for $513,690; the October model found more than five times as many novel LPEs "despite three months of hardening between the scans." In a sample of 166 exploits, OpenAI said the agents used 10 exploitation techniques that were new to it and to outside kernel-exploitation experts it consulted. "For safety reasons, we omit sharing the full technical details of these new exploitation techniques in this post," the company wrote. During the kernelCTF run the pipeline also found and exploited a bug in the QEMU emulator that ran the test environment; it had already been fixed upstream but not in the release build OpenAI used.
The post reported that when the scan and manual analysis finished after 10 days, 61 of the LPEs had already been patched upstream, 43 had staging patches and 299 were being reported and discussed publicly, which OpenAI credited to Linux maintainers and the wider security community. It did not say how many of the 948 OpenAI itself had reported, and it did not name the distribution. OpenAI wrote that, for the near and medium term, a low-privileged account on an operating system "that isn't up-to-date on its latest available patches should be considered functionally equivalent to root access for a well-resourced threat actor."
On November 5, OpenAI's GPT-6 Cyber announcement said "Our researchers used GPT-6 Cyber to develop 948 privilege escalation exploits in Linux" and to identify 81 previously unknown high-severity vulnerabilities in Firefox. It reported a 48.5% score on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and classified the model as Critical in cybersecurity, "up from High for GPT-5.6 Cyber." The model became available through the API and Codex to a select group of approved Daybreak Red customers, with wider Daybreak Red access promised later. TechCrunch, which called the model ChatGPT Cyber, reported that OpenAI said "thousands" of security professionals take part in Daybreak, and quoted OpenAI researcher Eric Wallace: "we are now allocating a significant amount of our frontier compute towards discovering and patching vulnerabilities in open source codebases, fully pro bono." Axios headlined its report "OpenAI's new cyber model found ways to attack nearly 1000 Linux flaws."
Why It Matters
The report puts a measured price on autonomous exploit development against one of the most widely deployed codebases in the world: about $1,279 per working root exploit, compared with five-figure competition bounties for the same class of bug. It also gives a dated comparison between two model generations on the same harness and target (166 LPEs in July, 948 in October), which is the kind of capability trend that frontier safety frameworks are meant to track. OpenAI's own conclusion was that "Within the next few years, we believe that if there are vulnerabilities in a given software project, frontier models will likely be able to discover them."
OpenAI first assigned the Critical cybersecurity level to Astra on September 1; GPT-6 Cyber, which OpenAI offers for exploit development and authorized security testing, now carries the same rating. Its release follows the gated pattern OpenAI set out then: access through vetted defender programs rather than general availability, with safety claims resting on OpenAI's own evaluations.
Several things are not known from the public record. All figures come from OpenAI; no outside party has published a count of the bugs OpenAI reported or the CVEs assigned to them. OpenAI withheld the new exploitation techniques. The time between OpenAI's reports and fixes reaching users of the unnamed distribution, which is what decides whether the exercise reduced risk, was not reported.
§ How to read the metadata
- Landmark
- Fundamentally alters the trajectory; 2–5 per year.
- Major
- Meaningfully shifts the landscape; 2–4 per month.
- Notable
- Worth documenting; significance can be upgraded later.
- Confidence
- High = primary sources corroborate. Medium = credible secondary only. Low = provisional. Disputed = credible sources disagree.
- Contestation
- Uncontested = no formal challenge. Contested = at least one challenge filed and not resolved by a correction. Corrected = a challenge was upheld and the entry corrected; see its revision history. Superseded = replaced by a later entry. Unresolved = dispute still open.
Significance history
Significance history
- NotableInitial rating at publication.
- MajorAutomated discovery of hundreds of kernel root exploits, released as a product, meaningfully shifts the cyber landscape.