Cyber-Permissive Model
An AI model specifically trained or fine-tuned to perform offensive and defensive cybersecurity tasks that standard consumer AI safety filters would normally block.
Most AI models sold to the public are built with guardrails that prevent them from helping users find security vulnerabilities, analyze malware, or generate exploit code. Those restrictions exist for obvious reasons. The same capability that helps a security researcher find a flaw in a system could help a criminal break into one.
A cyber-permissive model deliberately relaxes those restrictions for vetted users. The “permissive” in the name does not mean the model will do anything. It means it is permitted, within a controlled access program, to engage with cybersecurity tasks that standard models refuse. Think of it as a licensed version of an otherwise restricted capability, similar to how a licensed pharmacist can dispense drugs that are illegal on the street.
Anthropic’s Claude Mythos Preview (accessed through Project Glasswing) and OpenAI’s GPT-5.4-Cyber are the two most prominent examples as of early 2026. Both are restricted to vetted participants: government agencies, financial institutions, and security professionals who have passed a formal review process. Access is not public, and the application process is controlled by the lab, not by the purchasing organization.
The governance question these models raise is significant. Who decides who is vetted? What happens when a vetted organization misuses access? And as these models expand internationally, with Glasswing now extending to UK banks, which country’s rules govern what the model is permitted to do? These are not hypothetical concerns. They are the live policy debates happening right now inside the agencies and institutions deploying these tools.