One-Sentence Prompt Injection Flips 27% of AI Decisions

Patronus tested nine production decision models and found that a single injected sentence in customer free-text flipped 27% of correct decisions, with success rates reaching 80% after about 12 attempts, exposing serious prompt-injection risks in decision-making AI.

2026-11-20 ~ 2026-11-20 · 2 related posts