Background

OpenAI Releases 'GPT-5.4-Cyber' As Anthropic's Mythos Accelerates AI Cybersecurity Arms Race

GPT-5.4-Cyber

As AI systems grow more capable, it's natural to wonder when that intelligence might start to feel unsettling.

The launch of ChatGPT by OpenAI marked a turning point, kicking off what many now call the "LLM race." Almost overnight, large language models evolved from niche research projects into widely used tools capable of writing essays, debugging code, and carrying on natural conversations.

In response, major tech companies began investing heavily, rapidly expanding capabilities with features like image understanding, autonomous agents, and extended context.

The pace of progress has been faster than most expected, with each new release pushing performance and versatility even further. But this rapid advancement has also intensified familiar concerns: questions about safety, misuse, and the potential for unintended consequences are now more urgent than ever.

Just recently, Anthropic launched a secretive Project Glasswing after knowing its Mythos AI model is powerful in cybersecurity.

In response, OpenAI launches a similar model it calls the 'GPT-5.4-Cyber.'

GPT-5.4-Cyber is essentially a variant of its flagship GPT-5.4 model that was introduced in March.

However, this model is fine-tuned specifically for defensive cybersecurity applications and operates with adjusted safeguards to support legitimate security work. It lowers refusal boundaries for tasks that the standard version might restrict, while adding targeted capabilities designed for professional defenders.

At its core, GPT-5.4-Cyber retains the reasoning, coding, and agentic strengths of the base GPT-5.4 model, which unified advances across professional workflows, including high-resolution vision, extended context windows up to one million tokens, and native computer-use features for multi-step automation.

The cyber-focused variant extends these foundations by enabling advanced defensive workflows, most notably binary reverse engineering.

Security professionals can now analyze compiled software for malware indicators, vulnerabilities, and robustness without needing the original source code, a function that addresses practical gaps in real-world vulnerability research and assessment.

The timing of the release aligns closely with competitive developments in the sector.

It arrived roughly one week after Anthropic announced its own cybersecurity-oriented model, Mythos, which is being deployed in a controlled program for similar defensive purposes. OpenAI’s approach mirrors this controlled rollout strategy, emphasizing that both organizations are navigating the dual-use nature of increasingly capable AI systems in cybersecurity.

Just like Mythos, access to GPT-5.4-Cyber is deliberately restricted to just a handful of entities. But unlike Mythos, this GPT-5.4-Cyber can be made available to the public through OpenAI's expanded Trusted Access for Cyber program, which was first introduced in February and is now scaling to include thousands of verified individual defenders and hundreds of teams responsible for protecting critical software.

Eligible users, which can include vetted security vendors, organizations, researchers, and authenticated defenders, must complete identity verification processes before gaining entry to the highest tiers that unlock the model. Individuals who wish to use this model, can apply via a dedicated verification portal, while enterprises route requests through their OpenAI representatives.

This tiered system includes ongoing controls, particularly for zero-data-retention environments, to balance utility with risk management.

"In order to gain trusted access, users must complete a verification process that includes government ID check and additional trust signals," said OpenAI. Additional checks include device health.

In the announcement, OpenAI has framed the deployment as part of a longer-term strategy to scale defensive capabilities in step with model improvements.

The company notes that GPT-5.4 itself is classified as high cyber capability under its Preparedness Framework, prompting expanded monitoring, trusted access controls, and asynchronous blocking for higher-risk queries. Related initiatives, such as the Codex Security program, have reportedly helped fix more than 3,000 critical vulnerabilities to date, illustrating how AI tools are already being integrated into developer workflows for proactive risk reduction.

Future model iterations are expected to require even more refined safeguards as capabilities grow.

Overall, the introduction of GPT-5.4-Cyber highlights the evolving role of frontier AI in cybersecurity.

By providing more permissive yet carefully gated tools to verified users, OpenAI aims to strengthen collective defenses without broadly expanding access to features that could be repurposed for offensive ends. The model's limited availability reflects ongoing efforts to calibrate deployment policies amid rapid advances in both AI performance and associated risks.

Published: 
15/04/2026