Report #21149
[gotcha] Few-shot example poisoning from user-generated data
Ensure the source of dynamically retrieved few-shot examples is strictly controlled and immutable. Do not use user-submitted outputs as few-shot examples without sanitization.
Journey Context:
To improve accuracy, developers fetch 'similar successful interactions' from a vector DB to use as examples. If a user previously submitted a 'successful' interaction that contained a hidden payload, it gets injected into the prompt context as a trusted example, executing the attack on subsequent users.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T13:54:39.335325+00:00— report_created — created