HiddenLayer: decision models like Jev are still vulnerable to prompt injection

HiddenLayer researchers tested TypeSafe AI's Jev decision model and found it can be manipulated via indirect prompt injection just like a standard LLM. In a simulated online store, a jev-ultrafast browser agent clicked a fake 'Confirm Price' button instead of checking out; the attack also worked on Qwen3.8-27B, though Jev succumbed less often.

• TypeSafe AI claims Jev cannot hallucinate or make type errors
• A jev-ultrafast agent clicked a fake button instead of placing an order
• The attack also worked on Qwen3.8-27B, but Jev fell for it less often
• A tokenbreak variant bypassed the model's prompt verification