OSAFIS

📗 OSAFIS PromptLab
👢 An educational lab for exploring language model behavior and gaining hands-on experience with the security of AI systems.
⚠️ PromptLab includes 20 practical scenarios and 4 defensive architectures. You can compare how a model responds to the same input under different conditions, adjust the difficulty of challenges, follow each experiment step by step, and save your progress report.
⚠️ Scenarios cover prompt injection, indirect inputs, information retrieval, and tool interactions. The focus is on understanding how each scenario works and how defenses affect the outcome.
⚙️ PromptLab runs on your own computer with a local model. It is available for Windows, macOS, and Linux, and supports Persian, English, Chinese, Russian, and Arabic.
➡️ Download and installation guide
🔄 We welcome suggestions and bug reports to help improve future versions:
💎 report@osafis.org | 🔗 https://osafis.org
✅https://t.me/OSAFIS

📗Introducing OSAFIS - A Framework for Understanding AI Security Failures
⚙️ Today’s intelligent systems are more than a single model. They bring together data, memory, RAG, agents, tools, infrastructure, people, and different levels of authority. The question is not simply “What attack occurred?” We also need to identify what was supposed to be protected, who had which authority, which security contract was violated, how the failure spread through the system, and what evidence supports those findings.
❔We have been researching this problem since mid-2025,Our work led to a project called: OSAFIS (Open Security Assessment Framework for Intelligent Systems)
🖌 OSAFIS is an open research framework for analyzing and assessing the security of intelligent and autonomous systems. We aim to provide an analytical model for understanding the structure of security failures across the whole system - a framework for asking better questions about AI security.
📌 We have now published the initial version of the project and its documentation:
🔗https://osafis.org
✅https://t.me/OSAFIS

Channel created