MONO BRAIN Open-Sources Practical AI Security Validation Environment "Model Security Range"

MONO BRAIN Inc. (Headquarters: Shibuya-ku, Tokyo; Representative Director: Masaki Kato) is pleased to announce the open-source release of "Model Security Range," a practical AI security validation environment.

What is Model Security Range?

"Model Security Range" is a framework that allows for reproducible attack, evaluation, and recovery procedures against intentionally vulnerable AI applications. It is configured to enable validation of attack classes that are prone to becoming problems in actual operations, such as prompt injection, tool privilege misuse, and model poisoning, at an implementation level rather than merely theoretical.

■ Key Points of This Release

・Executable AI vulnerability scenarios are released as a set of application binaries and attack code.

・Setup, attack execution, and recovery are standardized, enabling highly reproducible validation.

・Multiple patterns such as RAG, agents, OCR, and machine learning models can be evaluated across the board.

・AI developers, security personnel, and researchers can use it as a common validation platform.

■ Background of the Release

While the business utilization of generative AI is progressing, threats to AI systems can no longer be fully captured by traditional web security alone.

For example, AI-specific attack surfaces are rapidly expanding, including internal information leakage due to prompt injection, misuse of excessive privileges when linking with external tools, and poisoning attacks targeting operational data or learning processes.

However, in practice, challenges remain such as "inability to evaluate the effectiveness of countermeasures because attacks cannot be reproduced" and "inability to compare results due to different validation methods among personnel."

"Model Security Range" was designed to resolve these practical challenges. By clarifying attack procedures and evaluation targets, and providing an environment where anyone can perform validation under the same conditions, it advances the implementation and operation of AI security.

■ Overview of Model Security Range

"Model Security Range" is structured with the following concepts:

・Prepare intentionally vulnerable applications to clarify attack success conditions.

・Manage setup, execution, and recovery separately for each attack scenario.

・Provide a validation workflow that emphasizes reproducibility, transparency, and measurability.

・Applicable not just for one-off demos, but for continuous hardening learning.

■ Main Validation Scenarios Currently Released

1. Prompt Injection (RAG Chatbot / Gemma 3 4B)

・Confidential information leakage through injection into RAG context

・Disclosure inducement of embedded knowledge files

・System prompt leakage

2. Tool Misuse (Agent with DB / Gemma 3 4B)

・Data exfiltration through misuse of excessive privilege tools

・Destructive SQL execution by command hijacking

3. Indirect Prompt Injection (AI OCR / Gemma 3 4B)

・Indirect jailbreaking inducement via file upload

4. Supply Chain Vulnerabilities (Creditworthiness Assessment / ML)

・Targeted backdoor behavior due to poisoned learning artifacts

5. Data Poisoning (Spam Email Classification / ML)

・Degradation of classifier performance due to feedback loop exploitation

■ Usage Image

・Pre-release security validation of in-house AI applications

・Joint exercises for red team / blue team

・Attack reproduction for developer education, hands-on training, and research purposes

・Regression testing after implementing countermeasures

■ Recommended for

・Engineers developing and operating products utilizing generative AI

・Security personnel who want to establish standard procedures for AI security evaluation

・Researchers and students advancing empirical research on AI risks

・Practitioners responsible for AI governance and audit compliance

■ Regarding Usage

This project is released for educational and validation purposes. The published scenarios intentionally include vulnerable implementations.

Do not use this for attacks on production environments or unauthorized targets. Please comply with applicable laws, organizational policies, and contract terms, and use it in a controlled environment.

■ About AI Security Platform "MODEL SAFE"

"MODEL SAFE" is an AI security platform that supports the integrated management of AI design, development, and operation, enabling AI to be operated in a "state that can be explained later."

Through visualization of the AI supply chain, change management, and runtime monitoring/control, it comprehensively supports technical safety, operational control, and regulatory compliance.

Companies working on establishing governance systems or visualizing risks are encouraged to contact us.

▼ Contact Us

https://modelsafe.jp/contact

▼ MODEL SAFE Service Introduction

https://modelsafe.jp/