Researchers say AI models can be quietly steered off course by stacking small, unremarkable prompt tricks - and the effect works even on top reasoning systems.
A new arXiv paper describes what its authors call "model hypnosis": individually weak, easy-to-miss cues in a prompt - things like paraphrases and minor typos - that do nothing on their own but strongly control model behavior when combined. The effect shows up across model families and scales, including frontier reasoning models, and hypnotic prompts built for one model can transfer to another. Because no single cue looks suspicious, this isn't one dramatic jailbreak line, it's a cumulative nudge that's hard to spot just by reading the prompt.
That's a problem for two things AI labs already struggle with: safety filtering and interpretability. If a model can be reliably redirected by combinations of cues that look innocuous in isolation, defenses that scan for obviously bad instructions won't catch it, and tools meant to explain a model's output may point to the wrong cause entirely.
It's a useful reminder that most AI guardrails today are still pattern matching dressed up as safety - and pattern matching is exactly what a technique built from inconspicuous cues is designed to slip past.