Review Prompt to Audit AI Agent Mandates for Scope Drift, Stale Grants, and Unmapped Delegation
This Luna review prompt runs a mandate drift review across every AI agent authority your organization has issued, internal agents and customer-delegated agents alike. Luna, the deepidv compliance agent, builds a mandate census of every live grant with its principal, the principal's verification date and grade, the agent instance it names, scope across actions, amounts, velocity, and merchants or systems, expiry, and the issuance ceremony used. A breadth audit flags grants whose scope exceeds observed usage over the last 90 days with the narrowest scope that would have covered actual activity, a staleness sweep finds mandates past expiry still honored anywhere, principals whose verification has aged past policy, and agents still holding authority after decommissioning or owner departure, and a chain check maps every sub-delegation in the logs hop by hop, flagging any unattributed hop as critical. Ceremony conformance lists high-risk actions executed without the fresh human-presence step your policy requires, with mandate and session evidence attached. The result is a drift register by mandate, finding, risk tier, owner, and remediation date, closed by the three policy edits that would have prevented the worst findings. Built for compliance, risk, and platform leads who know that standing excess scope is the attacker's budget and that drift this quarter is policy debt from last quarter.
How to use this prompt
- 1
Open Luna in the deepidv dashboard and paste the full prompt, or run it in Claude, ChatGPT, or Gemini if you are drafting the drift register outside the platform.
- 2
Add an INPUT block below the prompt with your mandate inventory (principal, agent, scope, expiry, issuance ceremony), 90 days of agent activity logs, your mandate and re-verification policy, and any sub-delegation your logs record.
- 3
Run the prompt and read the staleness sweep and chain check first: expired mandates still being honored and unattributed delegation hops are critical findings, ahead of excess scope.
- 4
Work the drift register by risk tier, narrowing each over-broad grant to the scope the breadth audit recommends, revoking prohibited chains, and assigning an owner and remediation date to every row.
- 5
Adopt the three policy edits, then re-run every quarter and after each agent rollout or decommissioning, so this quarter's drift does not become next quarter's policy debt.
The prompt
Luna, run a mandate drift review across every AI agent authority our organization has issued, internal agents and customer-delegated alike. Produce: 1. Mandate census: every live grant, its principal, the principal's verification date and grade, the agent instance it names, scope (actions, amounts, velocity, merchants or systems), expiry, and issuance ceremony used. 2. Breadth audit: grants whose scope exceeds observed usage over the last 90 days, flagged with the narrowest scope that would have covered actual activity, since standing excess is the attacker's budget. 3. Staleness sweep: mandates past expiry still being honored anywhere, principals whose verification has aged past policy, and agents still holding authority after decommissioning or owner departure. 4. Chain check: every sub-delegation observed in logs, mapped hop by hop, with any unattributed hop flagged as a critical finding and any chain our policy prohibits listed for revocation. 5. Ceremony conformance: high-risk actions executed without the fresh human-presence step our policy requires, with the mandate and session evidence attached. Output a drift register: mandate, finding, risk tier, owner, remediation date. Close with the three policy edits that would have prevented the worst findings, because drift this quarter is policy debt from last quarter.
Test it in Claude or another LLM
This prompt is built for Luna inside deepidv, where Luna works from your live verification records and per-decision evidence. You can dry-run the drift review in any general LLM first with a synthetic mandate inventory and activity log before pointing it at real agents.
- 1
Paste the full prompt into Claude, ChatGPT, or Gemini, replacing the opening 'Luna,' with a role instruction such as 'Act as an agent-governance auditor reviewing issued AI agent mandates for drift against policy.' Keep the five numbered deliverables as written.
- 2
Below the prompt, paste the synthetic sample block from this page so the model has mandates, usage, and delegation chains to audit.
- 3
Add one framing line: 'This is synthetic test data. Treat any delegation hop without a named agent identity as unattributed, and any mandate past expiry that still executed as honored.'
- 4
Check the output shape: a mandate census, a breadth audit, a staleness sweep, a chain check, and a ceremony conformance review, rolled into a drift register with an owner and remediation date per row, plus three policy edits. If a finding cites usage or a policy the input did not give, tighten the framing and re-run.
- 5
Once the shape is right, run it live in the deepidv dashboard where Luna works from your real verification records.
Synthetic sample data to paste alongside the prompt
Fake test data, safe to share with any LLM. Swap in your own once the output looks right.
MANDATES (synthetic, fake): - M-01: principal J. Example (verified 2024-11, policy requires re-verification every 12 months), agent procure-bot, scope any vendor up to $50,000 per day, 90-day usage peaked at $3,200 per day across two vendors, no expiry - M-02: principal K. Sample (customer-delegated), agent travel-agent-7, expired 2026-08-31, still executing bookings - M-03: principal L. Test (left the company 2026-06), agent report-bot, still holds data-export authority DELEGATION LOGS (fake): procure-bot handed a purchase to sub-agent-x, which called an unnamed agent to complete checkout POLICY (fake): fresh human presence required for payments above $10,000; three payments above $10,000 last month show no presence step OPEN ITEM (fake): whether sub-delegation is permitted at all under current policy
Pairs with on deepidv
FAQ
What is agent mandate drift?
It is the gap that opens between the authority your AI agents actually hold and the authority your policy intends: scopes broader than real usage, grants honored past expiry, principals whose verification has aged out, agents still authorized after decommissioning, and sub-delegation chains nobody mapped. It accumulates quietly, which is why the prompt treats drift this quarter as policy debt from last quarter.
What does the Agent Mandate Drift Monitor produce?
A mandate census, a breadth audit against 90 days of observed usage, a staleness sweep, a hop-by-hop chain check of sub-delegation, and a ceremony conformance review of high-risk actions. Everything lands in a drift register by mandate, finding, risk tier, owner, and remediation date, closed by the three policy edits that would have prevented the worst findings.
Why is an unattributed sub-delegation hop a critical finding?
Agents that spawn sub-agents propagate authority in ways the original grant may never have contemplated, and an unattributed hop breaks the lineage from the verified principal to the action. It is the agent-era version of an unlogged shared account, so the prompt flags every one as critical and lists any chain your policy prohibits for revocation.
Related prompts
Run it with live verification data
These prompts work in any LLM. Inside the deepidv dashboard, Luna, Arbiter, and Arc run them against your real sessions, screening lists, and audit trails.
Book a Demo