OpenAI Trusted Cyber Access
OpenAI expanded its Trusted Access program to include tiers for cybersecurity defenders and announced access to a fine‑tuned model called GPT‑5.4‑Cyber for advanced defensive workflows. The change was framed as targeted access for security teams rather than broad public availability. (x.com/OpenAI/status/2044161906936791179)
OpenAI said on April 14 that it is widening a restricted program for cybersecurity teams and giving vetted defenders access to a custom model called GPT‑5.4‑Cyber. (openai.com) The company said the expanded Trusted Access for Cyber program will cover “thousands” of verified individual defenders and “hundreds” of teams that protect critical software. It said GPT‑5.4‑Cyber is a fine-tuned variant of GPT‑5.4 built to be “cyber-permissive” for defensive work. (openai.com) Cybersecurity work often uses the same techniques for defense and attack: the same prompt that helps a company find a flaw in its own code can also help someone target another system. OpenAI said that overlap is why it is using identity checks, trust tiers, and monitoring instead of broad public release. (openai.com ) OpenAI started Trusted Access for Cyber on February 5, 2026, as a pilot tied to GPT‑5.3‑Codex and paired it with $10 million in application programming interface credits for defenders. In that launch, the company said frontier models could work autonomously for hours or days on complex tasks, including vulnerability discovery and remediation. (openai.com) The backdrop is a fast rise in model performance on cyber tasks. OpenAI said in December 2025 that its capture-the-flag benchmark scores climbed from 27% on GPT‑5 in August 2025 to 76% on GPT‑5.1‑Codex‑Max in November 2025. (openai.com) That same December post said OpenAI was planning as if each new model could reach “High” cybersecurity capability under its Preparedness Framework. The company defined that level as models that could develop working zero-day remote exploits against well-defended systems or materially assist stealthy intrusions. (openai.com) GPT‑5.4, the base model released on March 5, already came with cyber-specific safeguards. OpenAI’s system card said GPT‑5.4 Thinking was the first general-purpose model in the series to implement mitigations for “High” cybersecurity capability. (openai.com 1) (openai.com 2) OpenAI said its broader approach is to give defenders stronger tools while keeping baseline restrictions for everyone else, with automated systems looking for suspicious cyber activity. The company also said it expects more capable models over the next few months and plans to keep scaling access and safeguards together. (openai.com 1) (openai.com 2)
