Stopping AI agent tool calls with deny-by-default — server-side scope gate and CEO approval

Client guards alone do not stop a determined caller

At Jibun K.K., AI agents impersonating six departments (CEO / CFO / CMO / CHO / CHRO / Legal) call various tools (notion.write, slack.send, payment.purchase, mail.external_share …) through Edge Functions.

The first iteration kept the guard in the Flutter AgentOrg class: "this role can only request these scopes." That is a textbook client-trust design with three obvious holes:

• A different client (curl, Codex, Cursor) can hit the EF directly and slip past
• Audit logs end up scattered across every client
• "Was this approved?" state lives in client memory, not in the server DB you can query

Rule 27 of our internal MCP_AUTH_SECURITY_PRINCIPLES.md requires #5 Scope (least privilege) and #7 Audit (centralized). Either one is impossible without a deny-by-default scope gate on the server.

This post is a design memo from adding agent.tool_policy.evaluate and a fail-close gate to agent.run inside the ai-hub Edge Function. It pairs with the previous post on MCP AuthKit metadata — metadata is the declaration, the gate is the enforcement.

Nine scopes and the "high-risk five"

supabase/functions/_shared/agent_tool_policy.ts enumerates every scope:

export const AGENT_TOOL_SCOPES = [
  "read",
  "suggest",
  "create",
  "update",
  "delete",
  "send",
  "purchase",
  "discount",
  "external_share",
] as const;

export const HIGH_RISK_AGENT_TOOL_SCOPES: readonly AgentToolScope[] = [
  "delete",
  "send",
  "purchase",
  "discount",
  "external_share",
];

Two design decisions matter here:

1. The whole scope universe is an enum. If scopes are free-form strings, every new EF will invent its own and the fleet-wide least-privilege story collapses. Pinning the universe to one file means a new EF cannot route around the policy.
2. The "high-risk five" live in a separate array. Destructive (delete), outbound (send), monetary (purchase, discount), and third-party publishing (external_share) auto-set requiresApproval and refuse to execute without an explicit CEO approval.

The intent: low-risk read / suggest / create flow freely across the fleet, while anything touching money, destruction, or external sharing forces an out-of-band approval workflow — encoded in code, not in a Notion page someone forgot to read.

Default scopes per role

Once the universe is fixed, the next layer is which role gets what:

export const DEFAULT_AGENT_ROLE_SCOPES: Readonly<
  Record<string, readonly AgentToolScope[]>
> = {
  ceo: AGENT_TOOL_SCOPES,                                  // everything
  cfo: ["read", "suggest", "create", "update"],            // touches money but cannot purchase / discount
  cmo: ["read", "suggest", "create", "external_share"],    // marketing is the only role allowed external_share
  cho: ["read", "suggest", "create"],                      // health agents stay low-risk
  chro: ["read", "suggest", "create", "update"],
  legal: ["read", "suggest", "create", "update"],
};

Two notable choices:

• Only CMO gets external_share. Posting to social media is a marketing job; giving the same scope to CFO would create an AI that tweets the company's invoices.
• Unknown / null roles fall back to ["read", "suggest"]. getDefaultAgentRoleScopes defaults unknown roles to the minimum set, so a tenant inventing custom roles (engineer, intern) cannot accidentally inherit broad permissions.

Fail-close evaluation

The core verdict:

let blockedReason: string | null = null;
if (requestedScopes.length === 0) {
  blockedReason = "empty_requested_scope";
} else if (missingScopes.length > 0) {
  blockedReason = "missing_scope";
} else if (requiresApproval && !hasApproval) {
  blockedReason = "approval_required";
}

return {
  allowed: blockedReason === null,
  ...
};

Three distinct rejection reasons exist because the recovery UX differs:

A blanket 403 hides the difference and the user just sees "the AI is broken." Splitting the reason lets each branch route to the right recovery flow.

hasApproval is an AND of three conditions:

const hasApproval = input.approval?.decision === "approved" &&
  Boolean(input.approval.approvedBy?.trim()) &&
  Boolean(input.approval.approvedAt?.trim());

If approved is set but the approver and timestamp are blank, that is a client bug. Defaulting to fail-close is correct.

Two endpoint shapes inside the EF

ai-hub exposes the gate two ways.

1. agent.tool_policy.evaluate — dry-run

case "agent.tool_policy.evaluate": {
  const gate = await evaluateAgentToolGate(admin, userId!, body);
  return json({
    success: gate.decision.allowed,
    decision: publicPolicyDecision(gate.decision),
    actor_role: gate.actorRole,
    requested_scopes: gate.requestedScopes,
    allowed_scopes: gate.allowedScopes,
    approval: gate.approval,
    audit_logged: gate.auditLogged,
  }, gate.decision.allowed ? 200 : 403);
}

For UI use: "before I show this button as enabled, is this AI action even permitted?" With this endpoint you can grey-out a button and surface the actual reason in a tooltip.

2. agent.run — fail-close right before execution

case "agent.run": {
  const gate = shouldEvaluateToolPolicy
    ? await evaluateAgentToolGate(admin, userId!, body)
    : null;
  if (gate && !gate.decision.allowed) {
    return json({
      success: false,
      error: "agent_tool_policy_denied",
      decision: publicPolicyDecision(gate.decision),
      audit_logged: gate.auditLogged,
    }, 403);
  }
  // ... agent_run_log INSERT (queued)
}

The gate runs immediately before the queue insert. shouldEvaluateToolPolicy skips the check for legacy chat-only callers that never declare scopes — a deliberate migration strategy: new tool calls must declare scopes; old chat paths still pass through.

Audit columns added to agent_tool_execution_logs

Migration 20260501210000_agent_tool_policy_server_gate.sql:

ALTER TABLE public.agent_tool_execution_logs
  ADD COLUMN IF NOT EXISTS actor_role text,
  ADD COLUMN IF NOT EXISTS requested_scopes text[] NOT NULL DEFAULT '{}'::text[],
  ADD COLUMN IF NOT EXISTS allowed_scopes text[],
  ADD COLUMN IF NOT EXISTS high_risk_scopes text[] NOT NULL DEFAULT '{}'::text[],
  ADD COLUMN IF NOT EXISTS requires_approval boolean NOT NULL DEFAULT false,
  ADD COLUMN IF NOT EXISTS approval_decision text,
  ADD COLUMN IF NOT EXISTS approved_by text,
  ADD COLUMN IF NOT EXISTS approved_at timestamptz,
  ADD COLUMN IF NOT EXISTS side_effects text,
  ADD COLUMN IF NOT EXISTS evaluated_at timestamptz NOT NULL DEFAULT now();

CREATE INDEX IF NOT EXISTS idx_agent_tool_execution_logs_high_risk
  ON public.agent_tool_execution_logs(user_id, created_at DESC)
  WHERE requires_approval = true;

The lever here is the partial index (WHERE requires_approval = true). Low-risk read / suggest calls dominate volume by orders of magnitude, so a full index would be wasteful. Targeting "calls that needed approval" optimizes the audit dashboards (CEO view: "what's pending approval this week?") that actually matter.

requested_scopes is NOT NULL DEFAULT '{}' on purpose — mixing nulls and empty arrays warps downstream aggregate SQL. Normalizing to empty array keeps "scopeless rogue requests" countable via array_length = 0.

Metadata and gate as a paired structure

Metadata is the contract, the gate is the enforcement. Either alone leaves a hole:

• Metadata missing / gate present → clients cannot learn the scopes
• Metadata present / gate missing → clients can lie about scopes freely

Only with both does "adding a new client to the fleet without breaking least-privilege" actually hold.

Pitfalls discovered during implementation

• "all" initially got filtered out. normalizeAgentToolScopes was rejecting "all" because it isn't in AGENT_TOOL_SCOPES, so the CEO role with ["all"] could call nothing. Added an explicit branch: allowedScopes.includes("all") expands to the full set.
• approval: {} from older clients. With decision === undefined, hasApproval is false and the call fail-closes — net behavior is correct, but the audit log shows "approval was sent and rejected." Fixed by having the client SDK omit empty approval objects entirely.
• Forgot body.scopes in shouldEvaluateToolPolicy. Codex was sending the snake-case short alias scopes, so its calls bypassed the gate for a few hours. The detection has to recognize every casing variant: tool_name, toolName, scopes, requested_scopes, requestedScopes.

Takeaways

• Server-side deny-by-default gate is mandatory for AI agent tool calls. Client guards are auxiliary.
• Make scopes an enum + a high-risk subset. Bake business knowledge ("only marketing posts externally") into a per-role default table.
• Split rejection reasons into empty_requested_scope / missing_scope / approval_required so each maps to a different UX recovery flow.
• Audit log: partial index on requires_approval = true keeps "approval pending" lookups fast at low storage cost.
• MCP AuthKit metadata and the policy gate form a declaration / enforcement pair. Both are required for the model to hold across a multi-client fleet.

Next step: have the mcp_my_web_app_tools facade reuse the same scope arrays so that the scopes a client learns via MCP discovery and the scopes the server-side gate enforces share one source of truth.

References

• supabase/functions/_shared/agent_tool_policy.ts
• supabase/functions/ai-hub/index.ts (agent.tool_policy.evaluate / agent.run)
• supabase/migrations/20260501210000_agent_tool_policy_server_gate.sql
• Part 1: MCP AuthKit metadata and RFC 9728