OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks

OpenAI introduced a new AI model, GPT-6 Cyber, designed specifically to assist cybersecurity professionals. The model surpasses its general-purpose counterpart, GPT-6 Astra, in most cyber-related tasks. Recognizing the potential misuse of such a powerful tool, OpenAI is currently limiting access to the model for a limited number of approved security customers.

The plan to release a specialized cyber model based on GPT-6 was announced in September. As AI models become increasingly adept at identifying software flaws and transforming them into functional attacks, businesses are under growing pressure to bolster their cybersecurity measures. Gartner predicts that global AI cybersecurity spending will reach $51.3 billion this year, nearly double the $25.9 billion spent in the previous year.

OpenAI is heavily investing in this market, recognizing its potential for enterprise growth. This is OpenAI's fourth cybersecurity-focused model this year, following releases in April, June, and August. The company recently appointed Dali Rajic as its chief revenue officer, who brings a background in cybersecurity software. Both Google and Anthropic are vying for a share of the security market.

Anthropic's Glasswing program grants vetted users access to capabilities that might be blocked by its public models, while Google offers vetted defenders access to its Gemini 4 Argon model without cyber guardrails through its Fairwind Program. GPT-6 Cyber specializes in advanced vulnerability research and exploit development. Researchers can use the model to attempt breaching a server to understand the potential impact of a flaw.

The same techniques can be employed in malicious attacks, leading to restrictions that may also hinder legitimate security work. OpenAI trains its Cyber models to enhance specialized cybersecurity skills and minimize task refusals that could benefit attackers and defenders alike. GPT-6 Cyber was tested on Firefox, Mozilla's web browser, and the Linux kernel, the core software powering many servers.

Despite these systems being extensively inspected and hardened by developers and security researchers, AI models can still uncover novel vulnerabilities. Researchers used GPT-6 Cyber to identify 81 previously unknown, high-severity Firefox vulnerabilities and create functional exploits for 948 Linux kernel vulnerabilities, including 588 previously undetected flaws.

The Linux exploits could enable an attacker with normal local access to gain root privileges, the highest level of control over the system. OpenAI currently offers two tiers of access for trusted cybersecurity professionals through its Daybreak program: Daybreak Red enables advanced activities like simulating attacks and creating tools for red-teaming, while Daybreak Blue focuses on defensive tasks such as reviewing code, analyzing malware, and investigating incidents.

GPT-6 Cyber will initially be accessible to a select group of approved Daybreak Red customers, with expanded access to all Red customers in the coming weeks.