OpenAI expands cyber defense access to thousands, launches GPT-5.4-Cyber

OpenAI's expansion of its Trusted Access for Cyber program reflects a strategy that has been developing over several years. Since 2023, the company has supported defenders through its Cybersecurity Grant Program and has been evaluating its models' cyber capabilities. In 2025, it began including cyber-specific safeguards in model deployments, and earlier this year launched Codex Security to identify and fix vulnerabilities at scale. The new GPT-5.4-Cyber model is the latest step in this progression, following cyber-specific safety training that began with GPT-5.2 and expanded through GPT-5.3-Codex and GPT-5.4.

The company's approach is grounded in three principles: democratized access, iterative deployment, and ecosystem resilience. It has seen steady improvements in agentic coding with direct implications for cybersecurity. Codex Security, which launched in private beta six months ago, has contributed to over 3,000 critical and high fixed vulnerabilities since its recent launch. OpenAI says it has reached over 1,000 open source projects with Codex for Open Source, which provides free security scanning.

The stakes hinge on whether OpenAI can scale access to legitimate defenders while maintaining appropriate safeguards. The company acknowledges that more permissive models require more restrictive deployments. The long-term test is whether its current safeguards remain sufficient for future models, whose capabilities are expected to rapidly exceed today's best purpose-built models. Access to permissive models may face particular limitations around no-visibility uses like Zero-Data Retention, where OpenAI has less direct visibility into the user and purpose of the request.