How Can AI Agents Read Untrusted Sources Safely?

AI agents face security risks like prompt injection when they read from untrusted sources, access internal knowledge, and communicate externally. The Dual-LLM pattern is proposed as an architectural guardrail to mitigate these risks. This pattern uses a privileged LLM to orchestrate tasks and call tools, while a separate, quarantined LLM handles the untrusted data in isolation. A non-LLM controller manages the workflow between them, preventing the privileged model from being exposed to malicious instructions. While this significantly reduces the attack surface, the pattern's limitation is that it doesn't make the data itself trustworthy, as malicious content could still be passed to the end-user.