Report #62107
[gotcha] Poisoned few-shot examples overriding system instructions
Dynamically validate few-shot examples or ensure they come from a trusted, immutable source. Do not use user-generated outputs as few-shot examples without sanitization.
Journey Context:
If an application uses user-submitted successful queries as few-shot examples for future users, an attacker can submit a query that looks successful but contains a hidden instruction. When injected as a few-shot example, it acts as a persistent indirect prompt injection, affecting all subsequent users.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T10:44:01.140623+00:00— report_created — created