OpenAI's 'Trusted Access' Expansion Strategy for Cyber Defense and GPT-5.4 Cyber Release

Through Trusted Access for Cyber, OpenAI presented a direction that allows broader use of powerful cybersecurity models while preventing misuse through trust-based access control.

The core idea is to manage access based on who can use the model, rather than restricting the model's capability. This keeps features useful for defense intact while allowing high-risk tasks only for verified users.

• Identity-based access (KYC): Individuals receive access after identity verification, enterprises after organization-level approval, and researchers receive invitation-based advanced access.
• Tiered Access: Lower tiers provide only limited functionality, while higher tiers allow use of stronger models.
• Default safeguards and real-time monitoring: This includes training to refuse malicious requests, behavior-based detection (classifier), and monitoring of suspicious activity.

The prohibited areas are also clear. Actions such as data exfiltration, malware creation, and unauthorized testing are not permitted.

OpenAI views this approach not as simply strengthening restrictions, but as a way to spread defensive capability faster. The reasoning is that as more security experts use AI, vulnerability discovery speed, response time, and security automation can all improve.

From an industry perspective, this shows that AI security deployment methods may be shifting toward a model of 'expanding to verified users', rather than being either fully open or fully restricted. At the same time, the standard for model competition is broadening from simple performance to the ability to deploy safely.

Other developments announced alongside this include support for advanced cybersecurity tasks through GPT-5.3 Codex, the provision of a defense-specialized model through GPT-5.4 Cyber, and a support program offering $10M API credits.