Keep your agents under control!
Duration: 00:09:17
Speaker: Dani Vilardell
Type: CLS - Lightning (Academic track)
Expertise: No prior knowledge
Event: Devcon 8
Date: Nov 2026
Human oversight is a pillar of AI safety. It is often exercised by a human in the loop (HITL) who approves critical actions taken by an AI agent. AI agents’ increasing ability to bypass guardrails, though, raises the question of how to enforce HITL robustly. When a task is blocked by a permission they lack, agents will often grant it to themselves if they have the means.
In this talk, I’ll discuss the risks of current HITL deployments and present experiments on a popular agentic wallet platform, showing that an agent will raise its spending cap without asking the human when a task requires more money than the cap allows. I’ll then introduce inalienable HITL, a new primitive that guarantees an approval was issued by the human and not the agent, and present Countersign, an iOS and Android app we developed that uses off-the-shelf phones as trusted devices to harden existing HITL systems.