Protecting a Financial Agent

This guide demonstrates how to use ABS Core to govern a Customer Support Agent that has the power to issue refunds and process payments.

The Problem

A generative AI agent might be manipulated via prompt injection to issue a refund that exceeds its authority or violates AML (Anti-Money Laundering) policies.

The Solution: Bonded Execution

With ABS Core, we define a Policy Pack that acts as a financial firewall.

1. Define the Policy

We create a policy that checks the refund_amount and the user_tier.

// financial-gateway.js
{
  "rule": "max_refund_limit",
  "condition": "event.payload.amount > 500",
  "action": "DENY",
  "reason": "Refund limit exceeded for standard tier"
}

2. Configure the Risk Budget

The agent is assigned a Monthly Risk Budget of $5,000. Each refund subtracts from this budget. If the budget hit 0, the agent is automatically Slashed (halted).

3. Verification in the Ledger

Every blocked transaction is recorded with a unique trace_id.

{
  "trace_id": "trc_9a2f1b",
  "verdict": "DENY",
  "policy": "financial-gateway",
  "reason": "Risk Budget Depleted"
}

Results

• Prevented Liability: The company is protected from "hallucinated" refunds.
• Real-time Enforcement: Decisions are made in <20ms at the edge.
• Verifiable Compliance: Auditors can see exactly why every refund was allowed or denied.