AI was built to reason, improvise, and solve problems. Securing it is fraught with challenges when traditional cyber security methods and tools are applied. ACLs, firewall rules, protocol restrictions, and guardrails remain necessary, but are insufficient because AI relies on language to interpret goals, explain decisions, negotiate constraints, invoke tools, and delegate work. Human trust dynamics offer a complementary control, using language to establish expectations and observed behavior to determine how much authority an agent has earned.

Three familiar principles of human trust give this behavioral control its structure: limits on transferred trust, assessment against behavioral expectations, and evidence required for recovery. Confidence in a person does not automatically extend to everyone they rely on; similarly, trust in an agent should attenuate through delegation. Trust grows when observed behavior stays within an expected range and contracts when commitments are violated. Recovery requires repeated evidence of changed behavior before responsibility is restored. Applied to AI, these principles make trust a continuing assessment and provide behavioral evidence for technical controls to grant, restrict, or restore authority.

Human trust principles provide a practical way to decide how much responsibility an AI agent earns and keeps. We define what trustworthy behavior looks like for the agent’s role, observe what it actually does, and compare its actions with those expectations. Keeping commitments supports continued responsibility. Breaking them calls for closer oversight or reduced freedom. Responsibility is restored through repeated evidence of improvement.

Seven behavioral domains provide a practical structure for establishing an AI agent’s behavioral baseline. An anonymized assessment shows how tests revealed commitments that held and behavior that departed from expectations. The prompts, response excerpts, and assessment criteria connect those findings to decisions about how much trust and responsibility the agent should retain.

Attendees leave able to define behavioral commitments, apply the tests, interpret a trust profile, and connect findings to oversight and authority. The starting point is one consequential agent: define three commitments, test one under pressure, and compare its claims with its actions.