# OpenAI spent $1.2 million hacking Linux... and it was a bargain

## Executive summary

OpenAI demonstrated the capability of advanced AI models, specifically GPT6 Cyber, to systematically discover hundreds of unique root exploits in the Linux kernel master branch. The process utilized a multi-agent pipeline—including scout, analyst, triage, and exploit agents—to find 948 unique root exploits for a cost of $1,213,000. This highlights a critical acceleration in the cyber security arms race, suggesting that low-privileged accounts on outdated operating systems are functionally equivalent to root access.

## Key takeaways

- AI-Driven Exploit Discovery: OpenAI used a swarm of agents to tear through the latest Linux kernel master branch, resulting in 948 unique root exploits. The cost to achieve this was $1,213,000, equating to approximately $1,280 per exploit.
- The Multi-Agent Exploitation Pipeline: The process involved four stages: 1) Scout (carving targets, resulting in ~5,000 audit targets); 2) Analyst (scanning targets, finding 2,38 candidate bugs); 3) Triage agent (reproducing bugs in an isolated VM, confirming 1,724 bugs); and 4) Exploit agents (weaponizing exploits on hardened systems, yielding 1,169 working root exploits).
- Increased Efficiency and Power: Comparing earlier runs (GPT 5.6 Cyber found 166 LPEs for ~$3,100), the newer model found five times as many exploits for half the price, demonstrating rapid advancements in AI security capabilities.

## Technical details

- Local Privilege Escalation (LPE): LPE is the process of moving from a low-privilege foothold to a high-privilege (root) user. This is possible because operating systems, at their core, operate on binary logic (ones and zeros).
- AI Model Capabilities: GPT6 Cyber is described as a security-tuned variant of GPT6 Astra, scoring 48.5% on Exploit Gem (compared to 32.1% for GPT5.6 Cyber). The model's ability to exploit targets, even when Kernel Address Space Layout Randomization (KASLR) and Supervisor Mode Execution Protection (SMEP) are active, is notable.
- Security Best Practices: OpenAI advises that a low-privileged account on an operating system that is not up to date on its latest available patches should be considered functionally equivalent to root access.

## Practical implications

- Developers must prioritize keeping operating systems and dependencies updated to mitigate the risk of low-privilege escalation.
- The increasing power of AI in security necessitates robust, automated code review processes (e.g., using tools that run code in a sandbox) before deployment.
- The discovery of exploits is now a commodity, suggesting that security testing must move beyond simple vulnerability scanning to include advanced, AI-driven analysis.

## Topics

Linux Kernel, AI Security, Local Privilege Escalation (LPE), Software Vulnerability, Build Engineering, Greptile

Source: https://www.youtube.com/watch?v=pf5tXJZM1Rw
