You can't prompt your way out of prompt injection
Prompt injection in AI agents cannot be prevented with better system prompts, only with architectural controls. MailKite's inbox agent is designed assuming compromise, using ACLs, domain ownership ver…