Jev and other decision models can still be tricked with prompt injection
Never making type errors does not mean never making errors
Recent evaluations of Jev, a decision model from TypeSafe AI, reveal that it may still be susceptible to attacks despite claims that it cannot hallucinate and does not make type errors. TypeSafe AI's Jev, which outputs probabilities and confidence scores, can still be exploited through indirect prompt injection, according to a security firm.
HiddenLayer, a security business, tested both Jev and a traditional LLM, Qwen3.8-27B, to determine their resistance to manipulation. Results showed that while Jev succumbed less often to the attacks, both models were vulnerable to clickjacking via indirect prompt injection. By adding a sponsored ad to a simulated online store, researchers were able to trick both models into clicking on buttons and taking unintended actions.
This suggests that even decision models, which promise cost and performance advantages, should be approached with caution. The researchers argue that a model returning a decision remains a language model at its core and its decision translates into actions in your system.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
