GPT-5.4-Cyber access limited to ‘highest-tier’ users

GPT-5.4-Cyber access restricted to ‘highest-tier’ users

OpenAI has announced a new AI model called GPT-5.4-Cyber. Like Anthropic’s Claude Mythos model, this new ‘cyber-permissive’ GPT-5.4 variant is intended to defend against cyber threats rather than for public use.

OpenAI says its GPT-5.4-Cyber variant has been purpose-built to prepare for the release of more capable models later this year.

“As we prepare for increasingly advanced models from OpenAI to arrive in the coming months, we are fine-tuning our models specifically to provide capabilities for cyber defence, beginning today with a version of GPT-5.4 trained to operate in environments that allow cyber threats: GPT-5.4-Cyber.”

Access to GPT-5.4-Cyber is limited to ‘highest-tier’ users who are “prepared to work with OpenAI to authenticate as cybersecurity professionals”.

OpenAI says this is because GPT-5.4-Cyber is “specifically optimised to expand cybersecurity capabilities and has fewer functional restrictions”.

“This is a version of GPT-5.4 that lowers the refusal threshold for legitimate cybersecurity work and enables new capabilities for complex defensive workflows, including binary reverse-engineering capabilities that allow security professionals to analyse compiled software for potential malware, vulnerabilities, and the robustness of its security without needing access to its source code.

Because this model is more flexible, we are beginning with a limited, iterative deployment among vetted security solution providers, organisations, and researchers.”

GPT (Generative Pre-trained Transformer) is a family of advanced neural-network models created by OpenAI. Using transformer architecture, these systems can produce human-like text, hold conversations, write code and analyse information. GPT provides the foundation for popular chatbots such as ChatGPT.