OpenAI, the developer of ChatGPT, unveiled a cybersecurity-focused artificial intelligence (AI) model, "GPT-5.6 Cyber," and moved to revamp its AI support program for cybersecurity professionals.
OpenAI said on the 11th it will split "Daybreak," a program that provides vetted security experts with a high-performance cybersecurity AI model for defensive use, into "Daybreak Red" and "Daybreak Blue" by defense purpose.
OpenAI said, "Going forward, attackers will use AI to carry out cyberattacks at unprecedented speed and scale," adding, "The goal of this overhaul is to enable trusted defenders to use cutting-edge AI before attackers can deploy attack-focused AI capabilities at scale."
Daybreak Blue allows use of state-of-the-art models including "GPT-5.6 Sol" and applies safeguards tailored to approved defensive tasks. It is the default tier recommended for most security experts, supporting work such as vulnerability discovery, secure code review, malware analysis, incident response, and patch verification.
Daybreak Red is an access tier for organizations performing more specialized security work such as advanced vulnerability research, exploit development, and red-team activities, and provides GPT-5.6 Cyber, a cybersecurity-specialized model developed based on "GPT-5.6 Sol." The model is specially trained to reduce unnecessary "refusals" during security work and is optimized for detecting security vulnerabilities and producing mitigation patches for them.
Both access tiers are offered only to individuals and organizations performing preapproved security tasks. During use, controls such as identity verification, account security, monitoring, restrictions on approved use cases, and legal compliance checks apply.
According to a high-difficulty cybersecurity assessment conducted by OpenAI, GPT-5.6 Cyber completed 95% of all requests. Compared with the previous-generation model GPT-5.5 Cyber (57.3%), its ability to carry out cybersecurity tasks improved significantly. GPT-5.6 Sol completed only 1.5% of requests in the same assessment.
OpenAI said it discovered about 400 real-world software vulnerabilities using "GPT-5.6 Cyber." In particular, it identified two undisclosed vulnerabilities in "V8," the JavaScript engine used by Google's Chrome web browser, reported them to Google, and Google fixed the vulnerabilities after confirming them, the company said.
OpenAI will also expand the Daybreak Cyber Partner Program with major security partners including Accenture, Cisco, Cloudflare, CrowdStrike, IBM, and Unit 42 at Palo Alto Networks. Key partners can integrate OpenAI's cutting-edge cybersecurity models into their products, managed security services, and customer projects.
An OpenAI official said, "This will allow more organizations to leverage advanced cybersecurity capabilities without having to build and manage specialized cyber AI programs themselves."