An agent manipulated by prompt injection may attempt an unauthorized tool call. A policy engine places the authorization decision outside the agent’s reasoning: it checks the proposed action against written rules before execution and returns an allow or deny decision. section 3.8 section 4.8.8
The orchestration layer queries the engine before invoking a tool. Policies describe which roles may perform actions on resources, and the decision can account for the requester, tool and circumstances. A denied action is stopped before it reaches the network. section 4.8.5 section 3.8
External authorization avoids scattering checks through agent code. Embedded logic can suffice for simple scenarios but becomes harder to audit as rules grow. Platform-native controls centralize management, though multicloud deployments may require duplicate configuration. section 4.8.5
Claims tool authorization
A claims assistant uses Cerbos to govern tool access. Customers and adjusters can read claims they own; adjusters can update status on assigned claims below $10,000. Larger claims require a separate approval path. section 4.8.5
Listing 4.1 Cerbos policy for ClaimAssist tool access .
resourcePolicy: resource: "claims_tool" rules: - actions: ["read_claim", "get_status"] effect: EFFECT_ALLOW roles: ["customer", "adjuster"] condition: match: expr: request.resource.attr.claim_owner == request.principal.id - actions: ["update_status", "flag_for_review"] effect: EFFECT_ALLOW roles: ["adjuster"] condition: match: all: Of: - expr: request.resource.attr.assigned_adjuster == request.principal.id - expr: request.resource.attr.claim_amount < 10000
The policy listing shows how roles, actions and claim attributes combine in authorization rules. section 4.8.5
An allow decision does not establish that the agent was not manipulated, that the tool is trustworthy or that the overall workflow is safe. Pair authorization with agent identity, short-lived credentials and records of actions and accountability. section 4.8.5 section 3.8 section 4.8.8



