Prompt injection in regulated industries: healthcare and banking
aiprivacysecurity
Where hidden instructions create the most risk in healthcare and banking, and the controls leaders should expect before AI touches sensitive data.
In regulated sectors, a fooled AI can mean exposed records, wrong payments or compliance problems. The same prompt injection basics apply, but the stakes and the oversight expectations are higher.
Healthcare
Assistants may summarize referral letters, patient messages or outside documents. A hidden line in any of these could try to make the AI reveal notes or alter a summary.
- Keep AI away from records it doesn't need for the task.
- Require clinician review before anything enters a chart or goes to a patient.
- Isolate patient data, caches included.
See more on AI in healthcare.
Banking and finance
Invoices, claims and customer emails are routine inputs, and they're also natural hiding places. The classic example is a hidden instruction to forward documents or change payment details.
- Require human approval for payments and account changes.
- Verify payment or contact changes through a second channel, such as a phone call.
- Log AI actions for audit.
See more on AI in banking.
Shared principles
- Treat anything the AI reads as information, not orders.
- Grant only the access each job requires.
- Put people in the loop for sending, paying, deleting and sharing.
- Keep customer data separate. A 2025 Stanford audit found seven AI providers sharing a speed-up cache across customers.
Questions to ask vendors
- How is untrusted content separated from instructions?
- What actions need human approval?
- How is data isolated between clients?
No control is perfect, so layer them. Start with What is prompt injection? and our Responsible AI-by-Design Framework.