GPT-5.4-Cyber
2026-04-14 · Publisher : OpenAI · Domain : Artificial Intelligence
OpenAI élargit Trusted Access for Cyber à des milliers de défenseurs vérifiés et lance GPT-5.4-Cyber, une variante de GPT-5.4 plus permissive pour la sécurité défensive.
What changes
- Variante « cyber-permissive » de GPT-5.4 qui abaisse le seuil de refus pour les travaux légitimes de cybersécurité.
- Ajoute des capacités de rétro-ingénierie binaire pour analyser des logiciels compilés sans le code source.
- Accès par paliers, avec vérification d'identité automatisée pour les individus et partenariats pour les organisations.
- Déploiement limité et itératif vers des éditeurs, organisations et chercheurs vérifiés.
OpenAI expands Trusted Access for Cyber to thousands of verified defenders and launches GPT-5.4-Cyber, a more permissive variant of GPT-5.4 for defensive security. ### What changes - “Cyber-permissive” variant of GPT-5.4 that lowers the refusal threshold for legitimate cybersecurity work. - Adds binary reverse-engineering capabilities to analyze compiled software without the source code. - Tiered access, with automated identity verification for individuals and partnerships for organizations. - Limited, iterative rollout to verified vendors, organizations and researchers. ### Specifications - Based on GPT-5.4 (see the GPT-5.4 entry for context and modalities). - Availability: on request, in the highest tiers of Trusted Access for Cyber. - Possible limits for uses without visibility (zero retention) and for access through third-party platforms. - API identifier: not disclosed. ### Pricing Pricing not disclosed at launch. ### Benchmarks No numerical score published by the vendor in the announcement. ### Key takeaways The first OpenAI model in the “Cyber” series: fewer refusals than the base model and binary reverse engineering on top, reserved for verified defenders. It opens the 5.5-Cyber and 5.6-Cyber series.
Links
Tags : LLM, OpenAI, GPT-5, Cybersécurité, Accès restreint