OpenAI officially launched GPT-5.4-Cyber on April 20, 2026, marking a significant expansion of the company’s efforts to integrate artificial intelligence into global cybersecurity infrastructure. Alongside the model release, the organization announced a major expansion of its Trusted Access for Cyber (TAC) program, which provides specialized access to its most powerful models for vetted security researchers and defensive organizations.
GPT-5.4-Cyber is a specialized version of the GPT-5 architecture, specifically optimized for defensive security operations. According to OpenAI’s technical documentation, the model was trained on a curated dataset of secure coding practices, threat intelligence reports, and network traffic patterns. Key features include the ability to perform real-time log analysis, identify zero-day vulnerabilities in proprietary codebases, and suggest remediation steps for complex multi-stage attacks. OpenAI stated that the model demonstrates a 40 percent improvement in vulnerability detection accuracy compared to its predecessor, GPT-5-Cyber.
The expansion of the TAC program is designed to address the growing demand for AI-driven security tools among critical infrastructure providers and government agencies. Under the new guidelines, OpenAI will grant Tier 1 access to a broader range of cybersecurity firms and non-profit security research groups. This access includes higher rate limits and the ability to fine-tune GPT-5.4-Cyber on private, sensitive datasets within a secure, isolated environment. OpenAI confirmed that the TAC program now includes over 150 partner organizations, an increase from 45 at the start of the year.
Mira Murati, OpenAI’s Chief Technology Officer, stated that the release is part of a broader strategy to ensure that AI-driven defensive capabilities outpace offensive uses of the technology. The company emphasized that GPT-5.4-Cyber includes strict safety guardrails to prevent the model from being repurposed for malicious activities, such as generating exploit code or automating phishing campaigns. These guardrails are monitored by a dedicated internal team that conducts continuous stress testing on the model’s outputs.
The program expansion also includes a new Cyber Grant initiative, providing 50 million dollars in API credits to academic researchers focusing on AI-based defense. OpenAI’s Head of Cybersecurity, Matt Knight, noted that the goal is to foster a collaborative ecosystem where defensive tools are developed transparently. The company also announced partnerships with several major cloud service providers to integrate GPT-5.4-Cyber directly into their security operations centers, though specific financial terms of these agreements were not disclosed.