OpenAI drops GPT-5.4-Cyber for defensive security

The launch of GPT-5.4-Cyber marks the beginning of the "Vertical LLM" era, where general-purpose safety filters are peeled back for verified experts in high-stakes domains.

• –Direct response to Anthropic's Claude Mythos (Project Glasswing), signaling a specialized model arms race
• –Lowered refusal boundaries specifically for binary analysis and vulnerability research solve a major pain point for security researchers
• –Tiered verification at chatgpt.com/cyber introduces a new "Proof of Identity" friction for high-capability models
• –Strategic pivot from "one model for everyone" to domain-specific variants with granular access control
• –Clear signal that LLMs are graduating from coding assistants to deep infrastructure security tools