OpenAI Launches Specialized Cybersecurity Model with Limited Access

OpenAI has announced a new AI model called GPT-5.6-Cyber, specialized for cybersecurity professionals. The model is designed to support advanced security tasks such as vulnerability discovery and attack method validation, and is provided in a form accessible only to authorized defense-side specialists.

GPT-5.6-Cyber is based on GPT-5.6 Sol, a general-purpose model OpenAI released in June of this year, with additional specialized fine-tuning applied to the cybersecurity domain. Specifically, it enhances capabilities for highly specialized tasks such as discovering zero-day vulnerabilities (previously unknown security flaws that remain unpatched) and developing exploit chains that combine multiple attack techniques. Another notable feature is that it has been adjusted to reduce refusal responses compared to standard models when handling dual-use requests that could serve either defensive or offensive purposes.

In terms of performance, GPT-5.6-Cyber achieved a 95% completion rate on OpenAI's internal benchmark "Advanced Cybersecurity Completion Rate." This represents a significant improvement compared to the previous generation specialist model GPT-5.5-Cyber at 57.3% and the general GPT-5.6 Sol at 1.5%. According to OpenAI, this benchmark targets advanced scenarios including exploit chain development, authentication bypass, and privilege escalation. Eric Wallace, a researcher at OpenAI, also posted a comment on X (formerly Twitter) positioning this model as "the first major capability advancement effort aimed at highly advanced cybersecurity tasks like exploit development."

The distribution method is limited, and not all ChatGPT users or API consumers can use it freely. Access to GPT-5.6-Cyber requires review and approval through Daybreak Red, a new program announced by OpenAI on the same day. Daybreak Red is targeted at specialized teams performing advanced security work in authorized environments, such as vulnerability research, penetration testing, and red team exercises that test an organization's security posture. Meanwhile, Daybreak Blue, also announced simultaneously, provides access to a partially restricted version of GPT-5.6 Sol to a broader range of companies and does not require as rigorous a review as Red.

The pricing structure has also been disclosed. GPT-5.6-Cyber is priced at $12.50 per 1 million input tokens and $75 per 1 million output tokens. One token is the minimum unit of text that an AI model processes and roughly corresponds to less than one English word. Compared to GPT-5.6 Sol's pricing of $5 for input and $30 for output listed in the same pricing table, the Cyber model is more expensive. Pricing for long context is not listed in this pricing table, and access continues to require individual review and approval through Daybreak Red.

This initiative can be seen as one answer to the difficult question of how far AI companies should go in the security domain. An AI capable of acquiring advanced cyber attack techniques could become a tool in the hands of malicious actors depending on how it is used. OpenAI's decision to impose strict access restrictions limiting use to "approved defense-side specialists only" reflects an awareness of this risk. Going forward, key points to watch will be whether the review process maintains transparency and whether mechanisms to prevent misuse function effectively.