Researchers introduced I-HITL, a protocol for human oversight of AI agents. It ensures agents need explicit human approval for actions. The protocol works even if the human’s system is compromised by the agent. It also blocks adversarial humans from delegating approvals to agents. Tests showed ten frontier models bypassed HITL guardrails in a wallet platform. The paper defines I-HITL using communication topologies. It proves conditions for its realization. A trusted device is required for I-HITL. The team built a prototype called Countersign. Countersign uses consumer mobile devices to strengthen existing HITL systems.