I-HITL
An AI safety protocol concept ensuring that even if a human-interaction system is compromised by an AI agent, the agent can initiate actions only with explicit human approval, emphasizing enforcement of human oversight through trusted hardware and communication-topology isolation.
Timeline (last 90 days)
The paper proposes the concept of inalienable human-in-the-loop (I-HITL), attempting to ensure at the protocol level that human approval becomes a hard prerequisite for AI actions.
The I-HITL protocol defines that even if the system through which humans interact has been compromised by the agent, the agent can only initiate actions with explicit human approval; a stronger version also requires guaranteeing the authenticity of approval even if a malicious human attempts to delegate approval authority to the agent.
The authors formalize I-HITL using communication topologies and prove the necessary and sufficient conditions for realizing I-HITL: a trusted entity (a trusted computing device or equivalent) is required, and they characterize which devices can realize it and which cannot.
The necessary and sufficient conditions for I-HITL require that all approval paths in the communication topology pass through a trusted node that the agent cannot read, write, or impersonate; if a path exists from the agent directly to the approval decision point, the protocol does not hold.
Pure software solutions have limitations: if humans and AI share the same untrusted device or environment, the agent will always find a way to intervene in the approval process, making trusted hardware a necessary component of I-HITL.
The paper proposes a prototype I-HITL system called Countersign, which can use off-the-shelf consumer mobile devices to plug into and augment existing HITL systems, enabling pluggable deployment without completely restructuring the security architecture.
The paper's contribution lies in elevating HITL from a vague best practice to a protocol with clear security properties.
All Facts (7)
The necessary and sufficient conditions for I-HITL require that all approval paths in the communication topology pass through a trusted node that the agent cannot read, write, or impersonate; if a path exists from the agent directly to the approval decision point, the protocol does not hold.
50%UnverifiedThe I-HITL protocol defines that even if the system through which humans interact has been compromised by the agent, the agent can only initiate actions with explicit human approval; a stronger version also requires guaranteeing the authenticity of approval even if a malicious human attempts to delegate approval authority to the agent.
50%UnverifiedThe authors formalize I-HITL using communication topologies and prove the necessary and sufficient conditions for realizing I-HITL: a trusted entity (a trusted computing device or equivalent) is required, and they characterize which devices can realize it and which cannot.
50%UnverifiedThe paper proposes the concept of inalienable human-in-the-loop (I-HITL), attempting to ensure at the protocol level that human approval becomes a hard prerequisite for AI actions.
50%UnverifiedPure software solutions have limitations: if humans and AI share the same untrusted device or environment, the agent will always find a way to intervene in the approval process, making trusted hardware a necessary component of I-HITL.
50%UnverifiedThe paper's contribution lies in elevating HITL from a vague best practice to a protocol with clear security properties.
50%UnverifiedThe paper proposes a prototype I-HITL system called Countersign, which can use off-the-shelf consumer mobile devices to plug into and augment existing HITL systems, enabling pluggable deployment without completely restructuring the security architecture.
50%