Connect each agent action to delegated authority, policy checks, and evidence. Start with the situation your team needs to control.
Agents that make purchases and payments
Your agent selects goods and submits a payment on your behalf.
What goes wrong
Injected delegation can introduce an unauthorized agent. Scope creep can take an action beyond the authority you granted. A replay can repeat a recorded request.
Shipped controls
SURADAR checks the chain of custody for delegation. Per-request signing with replay protection and transaction binding tie authorization to the amount and payee. Cedar policy checks and human approval gate sensitive actions.
Illustrative scenario
Finance operations agents
A treasury agent prepares supplier payments and passes them to a payment executor.
What goes wrong
Scope creep can exceed delegated payment authority. A confused deputy can act on a payee outside the configured resource constraints. Step order violations can put execution before a required prerequisite.
Shipped controls
SURADAR binds the transaction amount and payee to the signed request. Cedar policy and human approval control payment authorization. Delegation checks enforce configured scopes and resource targets; prerequisite checks detect step order violations.
Illustrative scenario
Travel and expense booking agents
An assistant books travel within an approved budget and sends the purchase to a booking tool.
What goes wrong
Injected delegation can introduce an unauthorized booking agent. Scope creep can exceed the permitted action. Replay can repeat a booking request.
Shipped controls
SURADAR verifies the delegation chain and signs each request with replay protection. Transaction binding ties the amount and payee to authorization. Cedar policy and human approval can gate the purchase.
Coding agents and MCP tool access
A coding agent calls tools through the Model Context Protocol (MCP) while reading repository content and tool responses.
What goes wrong
Prompt injection can redirect tool use. Scope creep can exceed granted tool access. Step order violations can bypass a declared prerequisite.
Shipped controls
SURADAR provides a toolguard injection classifier and an MCP gateway. Delegation checks verify tool authority. Configured policy and human approval gate sensitive tool calls; prerequisite checks detect step order violations.
Agent memory and retrieval (RAG)
Your agent writes shared memory and uses retrieval-augmented generation (RAG) to bring stored context into later tasks.
What goes wrong
An unauthorized actor can attempt a memory write. Retrieved content can contain prompt injection that redirects a later tool call.
Shipped controls
SURADAR memory provenance uses signed writes and retrieval attestation to verify the source of stored context. The toolguard injection classifier checks for injection patterns at the tool boundary. Provenance establishes origin; it does not establish factual accuracy.
Audit and regulatory evidence
Your team needs evidence of agent authority and control decisions for internal review or an EU AI Act assessment.
What goes wrong
Injected delegation can obscure the authority behind an action. Expired delegation and replay can appear in observed activity.
Shipped controls
SURADAR evidence packages bring delegation and control records together for review. Verification-guarded records support inspection of the chain of custody and policy decisions. Evidence supports an assessment; it does not certify or guarantee compliance. Glyphzero does not certify compliance; your auditor does.
Controls across use cases
Chain of custody
Transaction binding
Policy and approvals
Memory provenance
Evidence
Bring your agent workflow.
Review the authority, controls, and evidence your use case needs.
