A 1-cent transfer can turn a bank's AI assistant into a phishing page (blue41.com)
Blue41 demonstrated indirect prompt injection against bunq's assistant: hostile instructions hide in an incoming payment's memo field, then get pulled into the model's context when you ask any question that fetches transactions.
The thread's reframe stuck: this isn't SQL injection. There's no query to sanitize. The right analogy is phishing — and phishing has no clean technical fix.
Sign in to save