2w ago · datadog-security-labs
Researchers demonstrated a novel attack technique called 'counterfeit policy' that manipulates TypeSafe AI's Jev model, a structured decision-making classifier used for security controls, by injecting a maliciously crafted policy into untrusted input. This technique successfully steers Jev's classification to bypass security checks, such as LangChain's tool-risk gating, by redefining risk criteria within attacker-controlled content. The attack significantly reduces detection rates, dropping recall from 82.90% to 3.88% in prompt injection detection, and enables unauthorized actions to be misclassified as low-risk, demonstrating a critical vulnerability in using such models as security judges without proper input sanitization.