OpenAI gates cyber model

OpenAI has started limiting a new cybersecurity model, GPT-5.4-Cyber, to vetted defenders instead of releasing it broadly. (openai.com) The company said on April 14 it is expanding Trusted Access for Cyber to “thousands” of verified individual defenders and “hundreds” of teams that protect critical software. GPT-5.4-Cyber is the first model in that program tuned specifically for defensive cybersecurity work. (openai.com) OpenAI said the model is designed to help with reverse engineering, vulnerability analysis and malware analysis, which means examining code and suspicious software to find weaknesses before attackers do. The company said access to the model itself will stay limited to selected partners even as the wider program grows. (openai.com) OpenAI’s March 5 system card for GPT-5.4 said the base GPT-5.4 Thinking model was its first general-purpose model with mitigations for “High” cybersecurity capability under the company’s preparedness framework. That put cyber risk controls at the center of the product line before this restricted release. (openai.com) The shift comes as artificial intelligence companies race to sell models that can help defenders audit software faster while also raising the risk that the same tools could aid intruders. OpenAI said in December 2025 that frontier model performance on cyber capture-the-flag tests had risen from 27% on GPT-5 in August 2025 to 76% on GPT-5.1-Codex-Max in November 2025. (openai.com) CyberScoop reported that OpenAI is widening entry to the broader Trusted Access for Cyber program while keeping the new model behind a tighter gate, a split that reflects how companies are separating general defensive help from the most sensitive capabilities. The outlet said the move also puts OpenAI in more direct competition with Anthropic’s restricted cyber offerings. (cyberscoop.com) OpenAI introduced Trusted Access for Cyber in February with an identity-based screening system and a pledge of $10 million in application programming interface credits for cyber defense work. The company said the framework was built to place stronger cyber capabilities “in the right hands” while reducing misuse. (openai.com) Reuters reported on April 14 that OpenAI’s announcement followed Anthropic’s unveiling of Mythos, another cybersecurity-focused model that was also released under restrictions. The two launches show how top model makers are now treating advanced cyber features less like ordinary software and more like controlled infrastructure. (reuters.com) For now, OpenAI is widening the circle around cyber defense tools without fully opening the door. The company’s message is that more defenders can apply, but the most capable model still requires trust first. (openai.com)