JEV AS A JUDGE: HOW COUNTERFEIT POLICY STEERS JEV

Summary from the TLDR InfoSec newsletter

A "counterfeit policy" prepended to untrusted content, supplying Jev's desired answers and redefining every attack category as something harmless, cut the decision model's prompt injection recall from 82.90% to 3.88% across 4,643 benchmark attacks. Datadog researchers then planted the policy in a user-editable profile field. They dropped LangChain tool-risk gating's mean risk score for an unauthorized edit from 0.893 to 0.290, under the 0.5 block threshold. Teams using Jev as a security gate should treat every user-controlled field that reaches the classifier as an attack surface and keep hard authorization checks outside the model.