Why We Can't Retrofit Old Security Principles Onto AI Agents (opens in new tab)

Traditional security relies on axioms like separating code from data, but LLM-based agents blur these lines by treating user prompts and untrusted external content as identical semantic inputs\. Dr\. Ilia Shumailov argues that current defenses are fundamentally flawed: adaptive attacks bypass standard guardrails with over 90% success, and existing red-teaming incentives often perpetuate vulnerabilities rather than fixing them\. This session presents a breakthrough alternative—deployment archi...