Jev as a judge: How counterfeit policy steers Jev
Datadog tests whether prompt injection can steer TypeSafe’s Jev and bypass LangChain tool-risk gating.
Datadog Security Labs evaluates TypeSafe AI’s Jev, a System One model that returns structured decisions with calibrated probabilities rather than free-form text. LangChain and Vercel use Jev for tool-risk gating, blocking calls the model judges risky or insufficiently authorized. TypeSafe’s own documentation says injected instructions, misleading framing, and text arguing for a classification can influence Jev. Datadog tests detection on 12,295 samples from four public prompt-injection datasets using Noul and Choice outputs, and tests whether attacker-controlled content can steer classifications and bypass that gate.
61