An agent may decide to call a tool after reading an attacker-controlled instruction. Authorization determines whether the requested action is permitted, independently of the agent’s reasoning. At runtime, an enforcement layer checks who is asking, which resource or tool is involved, and the circumstances of the request. section 4.8.5 section 4.8.8
For agent tools, an external policy engine lets the orchestration layer request an allow-or-deny decision before invocation. Policies live outside the model’s reasoning and can be audited independently. Platform-native controls centralize policy management but may require duplicated configuration across clouds. Checks embedded in orchestration code can work in simple cases, but become harder to audit as rules grow; checks left to the model’s reasoning are vulnerable to prompt injection. section 4.8.5 section 4.8.8
Claims tool access
A claims assistant uses a Cerbos policy to govern tool calls. Customers and adjusters can read claims they own. Adjusters can update the status of claims assigned to them when the claim amount is 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
Inspect the policy’s role, action, ownership, assignment, and claim-amount conditions. section 4.8.5
Authorization also belongs at the retrieval layer. Document-level controls restrict retrieval to content a user may see; they are not output filters. Azure AI Search and Elastic offer document-level controls, while Pinecone and Weaviate support configurable metadata filtering and tenant namespaces. Preserve permissions as sources enter or change in the index, and continue testing and monitoring retrieval behavior. section 5.4.3 section 5.2.4
For coding agents, intercept tool calls, shell commands, and file edits before execution. Rules can block dangerous operations or require human approval for sensitive actions. Authorization alone does not prove that an agent was not manipulated, that a tool is trustworthy, or that the overall workflow is safe. section 4.8.8 section 4.8.5





































