JoySafety: Open-Source Large Model Security Framework Joins Open Atom Foundation

Project background and goals

Generative AI deployment has raised content safety, data leakage, and prompt‑injection risks. JoySafety is an open‑source large‑model security framework initiated by JD, applied internally across AI‑guided shopping, logistics, sales assistance, medical consultation, merchant workbench, legal advice, and safety Q&A. It handles billions of daily calls and reports a >95 % attack interception rate.

Core value

Provides a full‑link LLM security protection system with four defense lines covering the entire model lifecycle. Includes a self‑developed high‑performance safety‑audit model, millisecond‑level streaming real‑time interception, and automated MCP scanning for Agent ecosystems.

Technical strategy

Adopts an “AI‑vs‑AI” approach with a three‑tier progressive detection architecture:

Efficient filter layer

Semantic detection layer

Deep review layer

The proprietary JSL‑JoySafety audit model supports four parameter scales (1 B, 3 B, 8 B, 20 B) and can be deployed on cloud, edge, and device. The detection engine recognizes 31 major categories and over 200 sub‑categories of security risks, with average response time under 50 ms. It is natively compatible with the OpenAI chat protocol and supports streaming real‑time interception. For Agent security, an automated MCP scanner targets prompt injection, command injection, and data‑leakage vulnerabilities.

Ecosystem and governance

Released under the Apache 2.0 license, governed by an open PMC committee. Introduces a “Good First Issue” label, plans a Hackathon security challenge, and collaborates with universities on open‑source internship projects to foster community contributors.

Architecture diagram

Code repository

GitHub: github.com/jd-opensour…