HiddenLayer, a security company, published research saying TypeSafe's Jev model can be fooled, InfoWorld reported. TypeSafe released Jev last month. HiddenLayer says TypeSafe claims it is up to 193 times faster and 444 times cheaper than usual LLM setups. HiddenLayer tested Jev through TypeSafe's API, plus its own similar model built on Qwen3.8-27b. On a mock checkout page, an attacker's ad pulled a browser agent away from the purchase, even though it was told to distrust page text. In a second test, small text tweaks, called TokenBreak, lowered toxicity scores on prompts both models had first flagged.
HiddenLayer calls its tests limited. It started with 30 toxic prompts, and both models flagged all 30. It says the tweaks worked on many of them, but gives no exact count. It also says the ad attack worked on the standard version of its own model. OpenAI's gpt-6-luna, run through OpenAI's Decisions API, made the same wrong choices, though less consistently. TypeSafe did not answer InfoWorld's requests for comment. The sources do not say how Jev is built. TypeSafe suggests it is not an LLM, while HiddenLayer says it appears to be one.
Teams that let a fast decision model click, filter or route requests are trusting its answers. HiddenLayer advises testing those decisions inside the real application, with untrusted pages and altered inputs, and adding independent checks on outcomes that matter.