Trusted access for the next era of cyber defense
Key Points
- GPT-5.4-Cyber for vetted defenders
- TAC expands with tiered verification
- Binary reverse-engineering enabled (limited rollout)
Summary
OpenAI is expanding Trusted Access for Cyber (TAC) and introducing GPT-5.4‑Cyber, a version of GPT‑5.4 fine‑tuned to be cyber‑permissive for defensive use. Access is tiered and requires identity verification; the model is being rolled out iteratively to vetted vendors, organizations, and researchers with additional controls (e.g., limits on Zero‑Data Retention and third‑party / no‑visibility uses). This release continues prior investments in Codex Security, grants, and open‑source scanning to accelerate defenders while maintaining safety and accountability.
Key Points
- TAC expanded with additional access tiers and automated identity verification; individuals may verify at chatgpt.com/cyber and enterprises request access via their OpenAI representative.
- GPT-5.4‑Cyber: lowered refusal boundaries for legitimate security workflows, supports advanced defensive capabilities including binary reverse engineering and vulnerability analysis; limited, vetted rollout to high‑trust users.
- Safety model: iterative deployment, stronger verification and trust signals, visibility requirements for permissive access, and restrictions on no‑visibility modes (ZDR) especially on third‑party platforms.
- Ecosystem support: Codex Security improvements (helped fix >3,000 critical/high findings), a $10M cybersecurity grant program, and continued support for open‑source security tools.
- Practical guidance for engineers: apply for TAC to access permissive models, expect stricter onboarding and monitoring for high‑risk capabilities, and integrate model outputs into CI/code review for continuous security feedback.