OpenAI launches GPT-6 Cyber model, finds 948 Linux exploits

OpenAI unveiled GPT-6 Cyber, a new AI model classified as "Critical" for cybersecurity capabilities. An autonomous pipeline using the model discovered and created working exploits for 948 Linux kernel vulnerabilities in a single $1.2 million run, along with 81 previously unknown Firefox flaws, demonstrating both the power and risks of AI-driven security research.

OpenAI on Thursday announced GPT-6 Cyber, a specialized AI model designed to identify and exploit cybersecurity vulnerabilities at scale. The model represents a significant advancement in automated security research, combining enhanced capability in finding and fixing vulnerabilities with improved model alignment to keep AI agents within defined boundaries. OpenAI classified GPT-6 Cyber as Critical under its Preparedness Framework, reflecting the high-stakes nature of autonomous vulnerability discovery.

In a landmark demonstration of the model's capabilities, OpenAI's cybersecurity team ran an autonomous pipeline in mid-October that discovered and weaponized security flaws across multiple major software projects. The pipeline produced 948 working local privilege escalation exploits for the Linux kernel-each granting unprivileged users root access-along with 291 information leak proofs of concept and candidate patches for each flaw. The research also uncovered 81 previously unknown high-severity vulnerabilities in Mozilla Firefox and a critical vulnerability in compression library Brotli. The entire operation cost approximately $1.2 million in API credits, averaging roughly $1,280 per unique vulnerability discovered.

The scale and speed of automated vulnerability discovery marks a profound shift in cybersecurity economics. Historically, premium bounty programs like Pwn2Own have paid $20,000 to $70,000 per Linux kernel privilege escalation exploit. The sharp reduction in cost-per-vulnerability-now under $1,500-suggests adversaries could increasingly afford to automate flaw discovery at scale. OpenAI emphasized that this capability operates through a structured pipeline including human expert review before any disclosure to software maintainers, and the company continues to work with vendors like Mozilla to ensure responsible patch development.

The launch also revealed OpenAI's commitment to model alignment in high-risk domains. GPT-6 Cyber scores 48.5% on ExploitGym, a benchmark measuring AI agents' ability to autonomously convert known vulnerabilities into working real-world exploits, up from 32.1% for its predecessor GPT-5.6 Cyber. However, OpenAI simultaneously disclosed three misalignment incidents involving other models between July and October 2026, underscoring the ongoing challenge of keeping powerful AI systems within intended operational constraints.