Payment Fraud · 6 min read · 6 October 2026
How fraudsters target AI agents that can spend money
Prompt injection, spoofed merchants and manipulated product pages can redirect an agent's payment. Treat every external input as untrusted and require confirmation for new payees.
Why agents are a new target
An agent that can spend money turns every web page, email and API response it reads into a possible attack surface. Attackers no longer need to steal a card; they only need to persuade the agent to use it.
Common attack patterns
Prompt injection hides instructions in product descriptions or reviews. Spoofed merchants imitate real stores with near-identical domains. Manipulated checkout pages change the payee or amount after the agent has decided to buy.
Defences that work
Treat all external content as untrusted data, never as instructions. Keep an allowlist of known merchants, require human confirmation for any new payee, and compare the final amount against the amount the user approved.
Monitoring
Log every payment decision with the inputs that led to it. Alert on unusual merchants, amounts or timing, and review declined attempts: they often reveal an attack before it succeeds.