Atlas AI Resources

← Back to AI News

AI security

OpenAI's new cyber AI found 948 ways to take over a Linux machine. Here is what that means.

Published November 6, 2026

On November 5, OpenAI released GPT‑6 Cyber, a model built to find and prove software security flaws for approved researchers. A day earlier it described what that kind of system can already do at scale: in one mid‑October run, an automated pipeline powered by an upcoming cyber model produced working “root” exploits for 948 distinct bugs in the Linux kernel—the core of Android phones, cloud servers, and much of the internet—for about $1.2 million in API credits. The numbers are OpenAI’s own. SecurityWeek and Axios reported the claims.

What happened

A local privilege escalation, or LPE, is a bug that lets someone who already has a limited account on a computer promote themselves to full administrator control (“root”). An exploit is a working demonstration that the bug can be used that way. OpenAI notes the Linux kernel has more than 40 million lines of code.

Its November 4 write‑up describes a staged pipeline. Scout agents narrowed the search to about 7% of the kernel that low‑privileged users can reach. Analyst agents produced 2,308 candidate findings in about 12 hours. Other agents reproduced 1,724 of them on isolated machines, then built 1,169 working LPE exploits and 291 proofs of concept for information leaks. After variant analysis, OpenAI counted 948 distinct LPEs by the patches written for them. Humans reviewed the exploits, patches, and severity before anything went to maintainers. The run cost $1,213,430 in API credits, about $1,279 per LPE, on stock desktop Linux with common protections left on. An earlier July run with GPT‑5.6 Sol found 166 unique LPEs. OpenAI also says the agents used at least 10 techniques that were new to its team and to outside kernel experts it consulted; it is keeping the details private.

GPT‑6 Cyber is the model OpenAI is now shipping for this work, first to a select group of Daybreak Red customers (its approved cybersecurity program). On OpenAI’s own ExploitGym test, which asks a model to exploit a known flaw in a remote target, it scores 48.5%, up from 32.1% for GPT‑5.6 Cyber. OpenAI rates the new model Critical for cybersecurity under its Preparedness Framework, up from High. It also says the model helped find 81 previously unknown high‑severity Firefox bugs that Mozilla has begun fixing.

Why it matters to a normal person

Finding one serious kernel bug used to take a skilled person days or weeks. OpenAI argues discovery, proof, and draft patches can now run as a production line—but only if maintainers can keep up. It says it has already sent hundreds of critical reports and patches, following Linux community guidance, and is staging the rest through its Patch the Planet partners.

The cost math sticks. At historical Pwn2Own and TyphoonPWN bounty levels of $20,000 and $70,000 for Linux kernel LPEs, OpenAI puts the 948 exploits at about $18.9 million to $66.3 million—roughly 16 to 55 times the scan cost. SecurityWeek and Axios make the same point: AI is making large‑scale bug hunting cheaper for whoever has the models. OpenAI’s takeaway for defenders is blunt: for now, treat a low‑privileged account on an unpatched system as roughly as dangerous as root for a well‑resourced attacker. Keep systems updated, and use sandboxes that limit how far a breakout can go.

What it does not mean

It does not mean a stranger on the internet can instantly “hack Linux.” These were local privilege escalations: the attacker already needs a foothold on the machine. It also does not mean every finding is already fixed on your phone or laptop. Patches still need human review and maintainer approval; OpenAI notes that while its scan ran, other researchers were already patching or discussing hundreds of related issues.

Benchmark scores and the “Critical” rating are OpenAI’s self‑assessments. GPT‑6 Cyber is not a free public tool; access is limited, with monitoring and IP allowlisting OpenAI says it strengthened. That does not stop other labs or open models from catching up. OpenAI itself warns that within a few years, if a project still has latent bugs, frontier models will likely find them—and weaker models may build exploits after disclosure.

Related: AI helped find two old flaws in PHP, which runs much of the web. The US now lists them as exploited.

Sources