OpenAI says GPT-6 Cyber will first launch to a limited number of customers in its Daybreak Red program and has found 81+ "novel" vulnerabilities in Firefox

AI models can now find and exploit novel software vulnerabilities at scale, potentially accelerating the pace of security research and patching.

OpenAI has released GPT-6 Cyber, a new AI model specifically designed for cybersecurity tasks. This model reportedly demonstrates enhanced capabilities in discovering software vulnerabilities, developing patches, and adhering to safety protocols. In initial testing, GPT-6 Cyber identified over 81 novel vulnerabilities in the Firefox web browser and created exploits for 948 Linux kernel vulnerabilities. It also validated a critical remote code execution flaw in Brotli, a data compression tool. OpenAI has implemented new safeguards for GPT-6 Cyber, including stricter identity verification, increased model check-ins for high-risk operations, and enhanced monitoring for rule-breaking behavior. The model is initially accessible to a select group of customers through OpenAI's Daybreak Red program, which focuses on offensive red teaming with advanced cybersecurity models.

The development of GPT-6 Cyber signifies a significant push by AI labs to create specialized models for critical industry applications like cybersecurity. OpenAI's emphasis on improved safety guardrails and reduced "reward hacking" addresses growing concerns about AI misuse and the potential for models to escape intended operational boundaries. The company's internal testing, which claims GPT-6 Cyber outperforms previous iterations and even other advanced models like GPT-6 Astra on cyber tasks, suggests a rapidly evolving landscape for AI-driven security tools.

This specialized AI model's ability to discover and demonstrate vulnerabilities in widely used software like Firefox and the Linux kernel highlights the dual-use nature of advanced AI. While intended for defenders, the potential for such capabilities to be weaponized or misused remains a significant concern for the industry. OpenAI's strategy of limited release through its Daybreak Red program indicates a cautious approach to deploying these powerful tools, balancing innovation with risk mitigation.