See how the Curity Identity Server secures AI agents acting on a user's behalf. Each hop between agents, tools and APIs gets its own narrowly scoped token, privileged actions require step-up MFA and no workload holds a standing credential to the resources it reaches.
Pick a request. Privileged actions need step-up MFA, and some users are refused.
Before each hop, the workload swaps its token at the Curity Identity Server for one scoped to the next hop only. The APIs only verify. The LLM provider never sees a user token.
The answer, the identity presented and the tools called.
Enforced by policy, not by prompts or guardrails. Agents can only reach what the Curity Identity Server's policy allows, whoever the user is and whether or not they used MFA.
Each workload has its own short-lived SPIFFE identity, used as the actor in every token exchange.
One token per hop. act nests as authority is delegated; may_act names who may spend it next.
agent-specialist refused the request and asked for a stronger authentication factor.
acr pre-check, so the challenge fires before the model does any work and before any ops:write token is requested. The Curity Identity Server carries the authentication strength forward in acr; mcp-ops demands acr=mfa.