Illia Polosukhin
Talks
Slides from recent talks, newest first.
-
Building Secure Personal Agents
Given increasing concerns about prompt injection in LLMs and other attacks on AI systems, I posit that the best solution is not in the LLM training itself but rather in the harness. By addressing risk at the whole system level and treating the harness as a full operating system, we can maximize utility and context while effectively treating every interface as an attack surface. I describe NEAR AI's approach with our IronClaw product.
Open slides