Researchers Find Prompt Injection Vulnerabilities in Non-LLM Decision Model Jev

·1 views

Security researchers at Check Point have published findings showing that Jev, a typed decision model rather than a traditional language model, is susceptible to prompt injection attacks. The discovery is notable because prompt injection has largely been associated with large language models, not structured decision systems. Check Point's research suggests that the vulnerability class may be broader than previously understood, extending beyond generative AI. The findings raise new concerns about the security assumptions made when deploying non-LLM AI systems in sensitive or automated decision-making contexts.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How to Safely Deliver Buyer Images While Preserving Unmodified Originals

A software architecture guide outlines best practices for delivering purchased images to buyers while keeping original files immutable and privately stored. The approach recommends serving only moderated, processed renditions by default, releasing originals solely when a buyer's order explicitly grants that entitlement. Every delivered image must be traceable to a specific source revision, crop specification, and moderation decision to prevent unreviewed content from reaching paying customers. Alerting systems should trigger on confirmed policy bypasses rather than every failed crop, and dashboards must identify failures by ratio and transform version to avoid masking localized errors. The guide also stresses that authorization decisions must precede byte-range handling, and that changing source files or crop algorithms silently invalidates prior moderation coverage.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

TypeSafe AI Launches Jev, a Model Built for Structured Decisions Over Text Generation

TypeSafe AI has released Jev, its first public 'System One' model, designed to make fast, structured decisions that software can consume directly rather than generating natural language responses. The model is inspired by Daniel Kahneman's System 1 and System 2 thinking framework, positioning Jev as a quick, intuitive decision-maker suited for backend applications. Instead of parsing generated text into booleans or scores, developers define typed questions and receive structured outputs such as yes/no answers, categorical choices, or scored evaluations in parallel. Jev uses a new architecture combining a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions (RLCD) to deliver calibrated probabilities alongside each decision. The model targets use cases like routing support tickets, flagging sensitive prompts, approving agent actions, and triaging transactions where structured outputs matter more than human-readable explanations.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tutorial: Build an Adaptive Python Tutoring API Using FastAPI and SQLite

A developer tutorial published on DEV Community walks readers through building PyMentor, a lightweight adaptive tutoring API for Python learners. The API, built with FastAPI and SQLite, accepts a learner's code submission along with their identifier and topic, then queries an OpenAI model for structured teaching feedback. It tracks topic mastery scores in a local SQLite database and returns validated responses to the client. The project is intentionally scoped — it does not execute submitted code or replace human instructors, focusing solely on repeatable feedback workflows. Developers need Python 3.10 or later, an OpenAI API key, and basic familiarity with Python, JSON, and HTTP requests to follow along.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tutorial: Enforce Beauty AR Device Policies in CI Before They Reach Users

A new developer tutorial demonstrates how to build a policy gate for Tencent RTC Beauty AR integrations that prevents expensive effects like GAN and 3D segmentation from running on low-end devices. The approach encodes rollout decisions — such as which effects are permitted per device tier and what fallback profiles apply — directly as versioned, testable code. A lightweight Node.js project structure validates profile configurations, detects cyclic or incomplete fallback chains, and exposes runtime states including applying, degraded, and blocked. The validation runs entirely within GitHub Actions CI without requiring a camera or any live Beauty AR asset. The tutorial positions device classification as an application-specific input, encouraging teams to base tier assignments on measurements from their own supported hardware rather than generic benchmarks.

0 comments Read more at DEV Community