21.04.2026
4 min
209
OpenAI has announced a new model, GPT-5.4-Cyber, designed specifically for cybersecurity professionals. It should help defenders find vulnerabilities faster, but at the same time raises questions about the risks of using such technologies by attackers.
The announcement of GPT-5.4-Cyber comes just weeks after competitor Anthropic introduced its new model, Mythos.
“The rapid adoption of AI will speed up the work of defenders — individuals or groups who are tasked with ensuring that information technology systems and networks, data and user identities are secure,” OpenAI stated.
Additionally, the AI company announced that it would expand its Trusted Access for Cyber (TAC) program to provide authentication to thousands of individual defenders and hundreds of defense teams who protect mission-critical software.
One of the inherent problems of all AI systems is that they are “dual-use”, as an attacker could take a legitimate version of AI technology and adapt it into a tool that they could utilize to further their own malicious goals. One key example of this is when attackers take a model that was developed using legitimate means to help find and fix weaknesses in widely used software, but instead, they use this same model to identify vulnerabilities in these same types of software prior to the patches being released and put into place. This puts users at risk because they have been exposed to additional potential attacks.
“By making our models available to the greatest number of people possible, we want to limit the potential misuse of our models, and also build stronger defenses through focused and incremental deployments,” OpenAI said.
“Our approach is to grow with the needs of cybersecurity and scale with the advancements in our model’s capabilities by increasing access to legitimate defenders, while continually enhancing the defenses,” Openai added.
ChatGPT developer says ChatGPT’s AI-powered application security agent called Codex Security has found over 3000 high-priority and heavily patched vulnerabilities.
OpenAI’s small-scale test came after Anthropic previewed its Mythos model, which is being rolled out under controlled conditions as part of the Glasswing project. According to Anthropic, Mythos has discovered “thousands” of vulnerabilities in operating systems, web browsers and other software.
