Jev Is Not a Language Model, but It Breaks Like One

Researchers tested Jev, a new AI model that returns typed decisions rather than text, and found it vulnerable to prompt injection attacks despite its different architecture from traditional language models. Every configuration tested could be manipulated to change its verdict at a cost of about 50 cents per successful attack, with reasoning being the only meaningful defense mechanism—which Jev lacks. The vulnerability is particularly concerning because these decisions are wired directly into systems without human review, unlike text outputs that people typically examine before acting on them.