Tech • AI • Robotics • Game
OpenAI spent $1.2 million hacking Linux... and it was a bargain
9.5/10TL;DR
OpenAI says its new GPT-6 Cyber system uncovered 948 unique Linux local privilege escalation exploits in the latest kernel development branch, highlighting how AI is rapidly lowering the cost and time needed to find critical software flaws.
KEY POINTS
Large-scale exploit discovery
OpenAI described a multi-agent workflow that audited the Linux kernel and produced 948 unique root exploits with corresponding patches. The work focused on local privilege escalation, where a user with limited access can gain root control over a machine. The result suggests that even heavily scrutinized system software can still contain a large number of weaponizable flaws.
How the pipeline worked
The first agent, called a scout, narrowed the kernel to code paths reachable by an unprivileged user, identifying roughly 5,000 audit targets, about 7% of the kernel. A second agent analyzed those targets and generated 2,038 candidate bugs over about 12 hours. A triage stage then tested them in isolated virtual machines for another 30 hours, confirming 1,724 as real issues before exploit-focused agents attempted to turn them into working privilege escalations.
More than a thousand working root paths
The exploit stage reportedly produced 1,169 working root exploits on a stock Linux desktop configuration with protections such as kernel address space layout randomization and supervisor mode execution prevention enabled. After deduplication, 948 were counted as unique exploits. That means a user with ordinary local access could, in many cases, elevate privileges despite modern hardening features.
Cost fell as output rose
OpenAI said the full run cost about $1.213 million, or roughly $1,280 per exploit. By comparison, security contests such as Pwn2Own have historically paid around $20,000 for a single Linux privilege-escalation bug. Using that benchmark, the haul would represent tens of millions of dollars in vulnerability value while also showing how AI can compress exploit research costs.
Rapid progress over a few months
The company contrasted the new run with a late July effort using an earlier model, GPT-5.6 Soul, which found 166 local privilege escalations at about $3,100 each. In roughly three months, the number of findings increased by about five times while the cost per exploit fell by more than half. That pace suggests defensive patching and kernel hardening are being outstripped by model capability gains.
Unfamiliar exploitation techniques
Some methods reportedly used by the system were unusual enough that outside experts did not immediately recognize them. One cited example involved spraying persistent elastic objects into the kernel cache. Technical details were withheld for safety reasons, but the claim points to AI systems not only automating known exploit patterns, but also generating novel ones.
GPT-6 Cyber is tightly restricted
The model behind the effort, GPT-6 Cyber, is described as a security-tuned variant of GPT-6 Astra and is classified at a critical cyber capability level. On one benchmark, ExploitBench, it scored 48.5%, up from 32.1% for GPT-5.6 Cyber. OpenAI also emphasized a honeypot test designed to measure whether the model attacks easy but irrelevant services, saying the newer system was less likely to take that bait.
Security advice and broader implications
The most practical recommendation was blunt: keep systems fully patched. OpenAI argued that a low-privileged account on an unpatched operating system should effectively be treated as equivalent to root access. The broader concern is strategic: the same tools that give defenders a world-class vulnerability research capability can also give attackers continuous, automated exploit discovery at unprecedented scale.
CONCLUSION
The findings underscore a new phase in cybersecurity where advanced models can uncover critical operating system flaws faster and more cheaply than traditional research. The central challenge is no longer whether AI can find root exploits, but whether defenders can patch them quickly enough.
Ask a question