The model
The OATH model: four moves that make an AI agent provable
Own, Approve, Tether, Hold to account. A security model built for agents, and how the Identity MCP, Gateway MCP and identity service carry it out.
Blog
Notes on knowing your agent.
Posts from the Oathera team, newest first.
Identity
Host binding: the agent identity primitive that stops cloning
Agents are trivial to copy, so a key alone cannot be an identity. Host binding ties an agent to where it runs, so a clone on a new host is refused.
Security
Beyond borrowed keys: proof of possession on every agent call
Borrowed keys make whoever holds them the agent. Proof of possession makes every call carry a fresh single-use signature, so a stolen token authorizes nothing.
Accountability
A named person for every agent: accountability in agent identity
Accountability is the part of agent identity borrowed keys drop. Oathera records a principal who approves each agent and a named sponsor who answers for it.
Enforcement
Checked every time it acts: continuous authorization for AI agents
Authorizing once at sign-in is not enough. The Gateway MCP checks a proof and asks for a signed decision on every call: no valid decision, no call.
Perspective
Is this really Agent X? Oathera and the Blueprint for the secure agentic enterprise
The Blueprint Alliance's reference architecture starts where we start: every agent is a distinct security identity. Underneath its four questions sits one more: is this really Agent X?