OpenAI introduces GPT-5.4-Cyber for cyber defense Apfelpatient

Following Anthropic's Claude Mythos, OpenAI is now also releasing a specialized AI model for cybersecurity. GPT-5.4-Cyber is optimized for defensive security research and is only accessible to verified security experts.

The arms race for AI-powered cybersecurity is intensifying. Just last week, Anthropic, with Project Glasswing and Claude Mythos, unveiled an AI model that has identified thousands of vulnerabilities in all major operating systems – with Apple as one of its founding partners. Now OpenAI is following suit: GPT-5.4-Cyber is a specially adapted version of the current GPT-5.4 model, designed specifically for defensive cybersecurity applications.

GPT-5.4-Cyber differs from the regular GPT-5.4 version in one crucial aspect: The model has been trained to reject fewer requests during legitimate security work. By default, AI models refuse responses that could be misused for attacks—for example, for analyzing malicious code or reverse-engineering software. This poses a problem for security researchers who do precisely that professionally. GPT-5.4-Cyber specifically lowers this threshold.

What the model can do

Among the new capabilities is binary reverse engineering: security experts can analyze compiled software for vulnerabilities, malware potential, and security robustness – without needing access to the source code. In practice, this is a key tool for threat analysis and incident response.

OpenAI emphasizes that GPT-5.4-Cyber serves as preparation for even more powerful models that are expected to follow in the coming months. The current model is a first step in testing defensive AI cybersecurity tools in practice before the next generation is deployed.

Strictly controlled access

Because GPT-5.4-Cyber deliberately has fewer restrictions than regular models, access is strictly limited. OpenAI requires verification through its “Trusted Access for Cyber” program. Individual security researchers can get verified via a dedicated page, while companies can request access through their OpenAI contact.

Initially, only selected security providers, organizations, and researchers will have access. OpenAI plans a phased rollout and intends to closely monitor the use of the model to prevent misuse.

Anthropic vs. OpenAI: Parallels and Differences

The parallels to Anthropic's approach are obvious: Both companies develop specialized AI models for cybersecurity, both restrict access to verified partners, and both see defensive AI as an urgent necessity before offensively usable models fall into the wrong hands.

The difference lies in the approach: Anthropic has built a broad coalition of technology companies with Project Glasswing – including Apple, Google, Microsoft, and Nvidia. OpenAI relies more heavily on its own ecosystem and Trusted Access program. For Apple users, this means that Apple's operating systems currently benefit directly from Anthropic's Mythos model, while OpenAI's GPT-5.4 Cyber initially operates independently of Apple.

Both initiatives demonstrate that AI-powered security research will have become one of the most important application areas for Large Language Models by 2026. (Image: Shutterstock / Photo For Everything)