What is the story about?
OpenAI is taking a measured route with its latest artificial intelligence model, opting against a broad release for now. Instead, the company has introduced GPT-5.4-Cyber in a limited rollout, granting access only to a group of verified cybersecurity professionals.
We, last week, reported that OpenAI is working on the cybersecurity tool and today they have finally unveiled it. The idea is simple: test the system in controlled, high-stakes environments before making it widely available.
The users, with access, are expected to push the model to its limits, attempting to uncover vulnerabilities, bypass safeguards, and identify how it could potentially be misused. The insights gathered will help OpenAI strengthen the system and refine its defences.
This comes just a few days after Anthropic introduces
Claude Mythos.
GPT-5.4-Cyber is a specialised adaptation of OpenAI’s GPT-5.4 large language model, fine-tuned specifically for cybersecurity tasks. Unlike general-purpose versions, this model is designed to engage more directly with complex and potentially sensitive security-related queries.
To make testing more effective, OpenAI has relaxed some of the usual guardrails that would typically restrict responses to risky prompts. This allows cybersecurity professionals to explore how the model behaves under adversarial conditions and assess whether it could be exploited by malicious actors.
The goal is not to make the system less secure, but to better understand its limits. By simulating real-world attack scenarios, OpenAI aims to improve the model’s resilience against jailbreaks and other forms of manipulation, while also enhancing its defensive capabilities.
Access to GPT-5.4-Cyber is currently restricted to participants in OpenAI’s Trusted Access for Cyber programme. This initiative is designed for vetted cybersecurity experts, researchers, and organisations working on defence and threat prevention.
These users are selected based on their expertise and are tasked with systematically testing the model. Their role includes identifying weaknesses, stress-testing safeguards, and providing detailed feedback that can be used to improve the system before any wider release.
This approach mirrors established practices in cybersecurity, where ethical hackers are invited to probe systems for flaws before they can be exploited in real-world attacks. In the context of AI, such controlled exposure is becoming increasingly important as models grow more capable and complex.
The introduction of GPT-5.4-Cyber comes amid intensifying competition between OpenAI and Anthropic, both of which are racing to establish leadership in AI-driven security.
Anthropic recently unveiled its Claude Mythos Preview model, positioning it as a next-generation system with advanced capabilities. The company has claimed that the model has already identified vulnerabilities across major operating systems and web browsers, highlighting its potential as a powerful security tool.
A key difference lies in their design philosophies. Claude Mythos is presented as an entirely new model, while GPT-5.4-Cyber is a fine-tuned version of an existing system, adapted specifically for cybersecurity use cases. OpenAI’s approach focuses on targeted improvements and faster iteration, whereas Anthropic is emphasising a more foundational leap.
Despite these differences, both efforts point to the same trend: cybersecurity is rapidly becoming an AI-driven domain. As attackers and defenders alike adopt increasingly sophisticated tools, the battle is shifting towards an AI-versus-AI dynamic, where resilience, adaptability, and early testing will be critical.
We, last week, reported that OpenAI is working on the cybersecurity tool and today they have finally unveiled it. The idea is simple: test the system in controlled, high-stakes environments before making it widely available.
The users, with access, are expected to push the model to its limits, attempting to uncover vulnerabilities, bypass safeguards, and identify how it could potentially be misused. The insights gathered will help OpenAI strengthen the system and refine its defences.
This comes just a few days after Anthropic introduces
What is GPT-5.4 Cyber?
GPT-5.4-Cyber is a specialised adaptation of OpenAI’s GPT-5.4 large language model, fine-tuned specifically for cybersecurity tasks. Unlike general-purpose versions, this model is designed to engage more directly with complex and potentially sensitive security-related queries.
To make testing more effective, OpenAI has relaxed some of the usual guardrails that would typically restrict responses to risky prompts. This allows cybersecurity professionals to explore how the model behaves under adversarial conditions and assess whether it could be exploited by malicious actors.
The goal is not to make the system less secure, but to better understand its limits. By simulating real-world attack scenarios, OpenAI aims to improve the model’s resilience against jailbreaks and other forms of manipulation, while also enhancing its defensive capabilities.
Who can use it?
Access to GPT-5.4-Cyber is currently restricted to participants in OpenAI’s Trusted Access for Cyber programme. This initiative is designed for vetted cybersecurity experts, researchers, and organisations working on defence and threat prevention.
These users are selected based on their expertise and are tasked with systematically testing the model. Their role includes identifying weaknesses, stress-testing safeguards, and providing detailed feedback that can be used to improve the system before any wider release.
This approach mirrors established practices in cybersecurity, where ethical hackers are invited to probe systems for flaws before they can be exploited in real-world attacks. In the context of AI, such controlled exposure is becoming increasingly important as models grow more capable and complex.
Claude Mythos vs GPT-5.4 Cyber
The introduction of GPT-5.4-Cyber comes amid intensifying competition between OpenAI and Anthropic, both of which are racing to establish leadership in AI-driven security.
Anthropic recently unveiled its Claude Mythos Preview model, positioning it as a next-generation system with advanced capabilities. The company has claimed that the model has already identified vulnerabilities across major operating systems and web browsers, highlighting its potential as a powerful security tool.
A key difference lies in their design philosophies. Claude Mythos is presented as an entirely new model, while GPT-5.4-Cyber is a fine-tuned version of an existing system, adapted specifically for cybersecurity use cases. OpenAI’s approach focuses on targeted improvements and faster iteration, whereas Anthropic is emphasising a more foundational leap.
Despite these differences, both efforts point to the same trend: cybersecurity is rapidly becoming an AI-driven domain. As attackers and defenders alike adopt increasingly sophisticated tools, the battle is shifting towards an AI-versus-AI dynamic, where resilience, adaptability, and early testing will be critical.
