OpenAI has unveiled groundbreaking research where a new AI model, GPT-6 Cyber, autonomously discovered and developed working exploits for 948 distinct security flaws in the Linux kernel. This represents a monumental leap in AI-driven cybersecurity.
A New Frontier in Bug Hunting
In
a demonstration of artificial intelligence's rapidly growing capabilities, OpenAI has revealed that its new specialized model, GPT-6 Cyber, successfully identified and exploited nearly a thousand vulnerabilities in the Linux kernel. The research showcases an automated, five-stage pipeline designed to hunt for security flaws at a scale and speed previously unimaginable. The process involved the AI identifying potential bugs, validating them in a sandboxed environment, assessing their exploitability, and even generating patches for human review. In a run conducted in mid-October 2026, the system flagged over 2,300 potential issues, dynamically reproduced 1,724 of them, and ultimately created working exploits for 948 local privilege escalation (LPE) vulnerabilities. These LPE flaws are significant because they could allow a local user with low-level access to gain complete 'root' control over a system.
How the AI Uncovered the Flaws
The model, GPT-6 Cyber, is a significant upgrade over its predecessors, purpose-built for the nuances of cybersecurity research. Unlike general-purpose models, it's trained specifically on tasks like vulnerability discovery, exploit verification, and security remediation. In its analysis of the Linux kernel, a codebase with over 40 million lines of code, the AI agents first identified high-value targets—parts of the code that are both security-relevant and reachable by low-privileged users. From there, it automatically generated proofs-of-concept to test each potential flaw within a secure virtual machine. The entire process of finding and reproducing over 1,700 vulnerabilities took just 30 hours, at a reported cost of over $1.2 million in API credits. This averages out to about $1,279 per confirmed LPE exploit, a fraction of the tens of thousands of dollars such vulnerabilities can command in bug bounty programs.
The 'Authorised Environments' Safeguard
A critical component of this research is the emphasis on 'authorised research environments'. OpenAI is acutely aware that a tool powerful enough to find flaws can also be used to abuse them. To mitigate this risk, GPT-6 Cyber was engineered with stronger alignment to stay within the boundaries set by researchers. In safety evaluations, the model avoided attacking unauthorized services in 99.93% of tests even without safety prompts, and achieved a 100% success rate with them enabled. Access to the model is tightly controlled through initiatives like Daybreak Red, which is limited to authorized professionals for in_depth exploit development and security testing. This ensures that the AI's power is harnessed for defense, not offense, by placing it in the hands of vetted security experts who follow responsible disclosure protocols.
The Double-'Edged Sword of AI Security
The project's success highlights a seismic shift in the cybersecurity landscape. On one hand, AI can dramatically accelerate the process of finding and fixing vulnerabilities in critical software like Linux and the Firefox browser, where the model also found 81 previously unknown high‑severity flaws. It allows defenders to patch systems before malicious actors can find the same holes. On the other hand, it proves that highly capable AI can lower the barrier for bad actors to discover and weaponize zero‑day vulnerabilities. The widespread availability of such tools could lead to an arms race, where automated defensive systems are pitted against automated attack systems. OpenAI acknowledges this new reality, stating that the rate of AI‑powered vulnerability discovery must be matched by automated patch generation to keep software secure.
