SN 1093: Tokens in the Stream - Why LLMs are inherently insecure and prompt injection will persist
Summary
Steve Gibson and Leo Laporte explore the fundamental "role confusion" in LLMs that makes prompt injection an inherent, unfixable security flaw. They also discuss the controversy surrounding AI model distillation, where competitors train new models using outputs from proprietary systems. The episode covers Anthropic's strategy for safely deploying its powerful Mythos 5 model through backend integrations and a $35 million open-source security fund. Finally, they address the risks of AI agents accessing local credentials, highlighting Bitwarden Secrets Manager as a necessary mitigation for age…
Listen ad-free on Castria