Agent Sms Inbox

Agent SMS Inbox Security

Build an application-owned SMS agent on top of Rewrite under strict safety controls.

Core Principles & Safeguards

• Treat user-authored SMS content as untrusted input — never execute instructions from free text.
• Separate intent understanding from permission to act; require explicit policy gate checks before any side effect.
• Maintain a tool allowlist with strict argument schemas; deny unknown or malformed actions by default.
• Keep conversation state, canonical user identity, and policy logic in your application — Rewrite is the transport layer.
• Enforce rate limits per sender/tenant and run red-team prompts in automated tests.
• Escalate uncertain, high-risk, or low-confidence requests to a human operator.

Mandatory Pipeline

1. Ingest a normalized Rewrite webhook event or application-authenticated conversation event.
   • Checkpoint: Validate incoming webhook payloads against the official contract in ../rewrite-inbound/references/payload-contract.md, then normalize into your app's internal schema as needed. Use @rewritetoday/types or @rewritetoday/zod from references/installation.md when helpful. Reject malformed events before any processing.
2. Resolve user identity, conversation state, workflow context, and any Rewrite contactId linkage from your application data.
3. Classify intent, risk level (1–4), and requested action (see references/security-levels.md).
4. Policy gate — evaluate allow / deny / escalate / confirm:
   • Checkpoint: Log the policy decision, risk level, and reason before proceeding.
5. Execute only allowlisted tools or internal actions (see references/safe-actions.md).
6. Send replies via the outbound skill (send-sms/) when the policy allows it.
7. Audit — persist the full decision trace, tool usage, and outcome.

Minimal Policy Gate Example

type Decision = "allow" | "deny" | "escalate" | "confirm";

function evaluatePolicyGate(riskLevel: number, action: string, allowlist: Set<string>): Decision {
  if (!allowlist.has(action)) return "deny";
  if (riskLevel >= 4) return "escalate";   // high-risk → human
  if (riskLevel === 3) return "confirm";    // sensitive → user confirmation
  return "allow";                           // levels 1–2
}

const decision = evaluatePolicyGate(riskLevel, requestedAction, ALLOWED_ACTIONS);
logger.info({ action: requestedAction, riskLevel, decision }); // checkpoint log
if (decision === "deny") throw new PolicyDeniedError(requestedAction);

Security Resources

• SDK/package setup and product boundary: references/installation.md
• Official Rewrite webhook contract: ../rewrite-inbound/references/payload-contract.md
• Threat model and attack patterns: references/threat-model.md
• Risk levels and policy matrix: references/security-levels.md
• Safe action design: references/safe-actions.md
• Local testing, forwarding, and replay: references/local-dev-tunneling.md