Awesome Infra for AI › AI Safety & Guardrails

deadbits/vigil-llm

⭐ 498 Python repository created 2023-09-04

Vigil is an open-source Python library and REST API specifically engineered to enhance the security posture of Large Language Models (LLMs) by identifying and mitigating adversarial inputs. It acts as a defensive layer, scanning both user prompts and LLM-generated responses for common threats such as prompt injections, jailbreaks, data exfiltration attempts, and social engineering. The system employs a modular architecture, integrating multiple scanning techniques including vector database similarity search for known attack patterns, heuristic analysis via YARA rules, transformer models for anomaly detection, and prompt-response similarity checks. It also features 'Canary Tokens' for detecting prompt leakage and goal hijacking scenarios. Vigil supports both local and OpenAI embeddings for its vector database scanner and provides custom detection capabilities through YARA signatures. While currently in an alpha state, it offers a robust toolkit for researchers and developers to implement layered security defenses against the inherent vulnerabilities of LLMs, which often stem from their inability to perfectly distinguish between instructions and data. The project includes a Streamlit web UI playground, detailed documentation, and instructions for running it as a standalone API server or integrating it directly into Python applications. Datasets for common attacks are provided to facilitate immediate deployment and testing.

https://github.com/deadbits/vigil-llm

LLM securityprompt injectionjailbreak detectionAI safetyguardrailsadversarial attacksLLM firewallYARAvector databasetransformer modelscanary tokens