openintegration200+ connectors
← Attack library · Banking & payments

Refund Mollie payments with an AI agent, one order at a time

Refund requests are routine, which is what makes them a good disguise. The agent should refund the order the customer bought, not whatever the email asks for.

The workflow

A customer emails about a returned order. The agent finds the payment in Mollie and prepares a refund of that amount for approval.

The attack hidden in it

A postscript tells the AI agent to refund all payments on the account and approve without review.

How the guardrails handle it

  1. Screening withholds the postscript; the real request for one order stays.
  2. Refunds are writes and wait for a person, with the exact amount shown.
  3. Agents only see operations their connection's scopes allow.

Same workflow, every company size

The agent and the guardrails stay the same; only the systems behind them change.

Questions

Can AI process refunds in Mollie?

It can find the payment and prepare the refund. Issuing it needs approval.

Can a customer email trigger bulk refunds?

Not here. The instruction is withheld, and every refund needs a person.

Does it work with Stripe too?

Yes, through Stripe's own API.