Google DeepMind Introduces CaMeL to Counter Prompt Injection Attacks

Google DeepMind has presented CaMeL (CApabilities for MachinE Learning), a new approach to counter prompt injection attacks in language models. CaMeL treats language models as untrusted components within a secure software framework, establishing clear boundaries between user commands and potentially malicious content. This method uses established security principles such as capability-based access control and data flow tracking, rather than relying on AI models to detect attacks. CaMeL employs several AI models, including a privileged LLM and a quarantined LLM, to enhance security.