Kovant CEO Ali Sarrafi explains why agents are not just advanced chatbots, but autonomous workers that can be manipulated, tricked and prompt-injected if not properly secured.
Sarrafi argues that enterprises must stop trusting the LLM to police itself. Instead, he advocates for treating agents as “digital employees” that require strict onboarding, limited access rights, and external guardrails.
He details why security policies must sit outside the AI, using deterministic software to control non-deterministic agent behavior, and warns that without these layers, deploying agentic AI is like “giving a five-year-old your bank account.”
If you are a CIO under pressure to adopt autonomous agents but need a practical framework for governance, access control, and risk mitigation, this interview provides the blueprint for safe deployment.
