Jev as a judge: How counterfeit policy steers Jev

Summary

TypeSafe AI introduced Jev on September 15 as the first of its System One Models, which return structured decisions in parallel instead of generating text token by token. This post evaluates Jev's prompt injection detection and tests whether attacker-controlled content can steer its classifications to bypass LangChain's tool-risk gating.

Related items

• MediumLangflow's Critical Authenticated RCE Flaw Targeted in 1,423 Honeypot Exploit AttemptsSame vendor · The Hacker News
• MediumCVE-2026-108575: BerriAI LiteLLM improper authorization in secret resolution via api_keySame vendor · NVD/CVE Database
• HighCVE-2026-106119: LangChain MongoDBChatMessageHistory query injection via session identifierSame vendor · NVD/CVE Database
• LowGHSA-5x6v-p487-7qh2: LangChain: RediSearch Filter Injection via Unescaped Tag/Text ValuesSame vendor · GitHub Advisory Database
• HighGHSA-fvww-7h3r-vfhp: LangGraph SDK custom auth silently ignores actions= on resource decoratorsSame vendor · GitHub Advisory Database