OpenAI AI Finds 948 Linux Kernel Root Flaws
OpenAI says its AI pipeline found 948 Linux kernel flaws that let an ordinary user gain root access, with working exploits and patches, for about $1.2 million.
OpenAI says its AI pipeline found 948 ways for an ordinary Linux user to gain root access, and it built a working exploit and a candidate patch for every one of them. The company published the results on November 4 in a research post called "A thousand fewer paths to root". The whole run cost about $1.2 million in API credits, and it shows how fast AI is changing the work of finding and fixing security flaws.
How the pipeline found 948 flaws
OpenAI built a five-stage pipeline: discovery, dynamic validation, exploitability assessment, patch generation, and human review. An upcoming cyber model powered the system, and researchers checked each exploit, patch, and severity rating before anything was reported. Scout agents first picked 5,034 audit targets covering about 2.8 million lines of code, roughly 7 percent of the Linux kernel. Analyst agents then scanned those targets and produced 2,308 candidate findings within 12 hours. A triage agent tested each candidate in a virtual machine running a kernel with Kernel Address Sanitizer, a tool that catches memory errors. Within 30 hours it reproduced 1,724 real vulnerabilities, about 74.7 percent of the candidates. Exploit agents then tested the flaws against the protections a standard Linux distribution ships with, including SMEP, SMAP, KASLR, and HARDENED_USERCOPY. They produced 1,169 working local privilege escalation exploits, and variant analysis grouped those into 948 distinct flaws, each with its own patch.
What the run cost
The full run cost $1,213,430 in API credits. The exploitation stage was the biggest part at $775,929, and a successful exploit cost $359 on average. OpenAI calls these figures an upper bound, since better engineering could bring the cost down. For comparison, a July 2026 run using GPT-5.6 Sol found and exploited 166 flaws for $513,690. The newer model found more than five times as many in October, even though the kernel had three months of extra hardening in between. OpenAI also compared the run with bug bounty rewards. At past Pwn2Own and TyphoonPWN prices of $20,000 and $70,000 per flaw, the 948 exploits would be worth $18.9 million to $66.3 million, or 16 to 55 times what the run cost.
New tricks the AI discovered
OpenAI sampled 166 of the exploits and found the agents used 10 exploitation techniques that outside kernel experts had not seen before. The new tricks include fresh ways to spray objects inside the kernel cache, page grooming methods, data-only paths to root, and an arbitrary type confusion. OpenAI shared the techniques with third-party experts, who confirmed they were new, but it withheld the full technical details for safety reasons. The post is also honest about the pipeline's limits. In an earlier run, 20 of 200 reported exploits reached root through a different flaw than the one assigned. OpenAI added checks to confirm each final exploit used its intended flaw.
Fixes are already landing
The scan ran on the latest Linux kernel in mid-October 2026. By the time the scan and manual review finished 10 days later, 61 of the flaws had already been patched upstream, 43 had patches in staging, and 299 were being reported and discussed in public. OpenAI says it has sent hundreds of reports and patches to maintainers, starting with the most critical flaws, and is working with partners in its Patch the Planet program on the rest. The patches are candidates for maintainer review, not fixes already merged into the kernel. OpenAI's guidance for defenders is blunt: an unpatched system with a low-privileged account should be treated as equivalent to root access for a well-resourced attacker. The company recommends patching fast, rescanning code as models improve, and sandboxing with tools like Firecracker and gVisor.
Why it matters
The Linux kernel runs on Android phones, cloud servers, and supercomputers used by billions of people, and it has more than 40 million lines of code. Finding flaws in it used to take a separate, sustained investigation for each bug. OpenAI's run shows that discovery can now be run as a systematic process, with validation and patching built in. That cuts both ways. Defenders can now test their systems far more cheaply, but attackers with the same tools can find flaws faster too. The race is no longer only about who finds bugs first, but about who patches them first.
Sources
- OpenAI - https://openai.com/index/a-thousand-fewer-paths-to-root/ - GBHackers - https://gbhackers.com/openais-new-cyber-ai-model-finds-948-linux-privilege-escalation-vulnerabilities/ - StartupFortune - https://startupfortune.com/openais-ai-pipeline-turned-2308-linux-kernel-bugs-into-948-root-exploits/