Vigil

// Overview

Vigil is a security scanner for large language model prompts and responses. You give it the text going into or coming out of an LLM, and it runs that text through a set of modular scanners to flag prompt injections, jailbreaks, and other risky inputs. It ships as both a Python library you import into your own app and a REST API server you can run as a standalone service.

It is aimed at engineers building LLM-backed applications who want a detection layer in front of their model. Vigil bundles the detection signatures and datasets you need to self-host, and lets you add your own YARA rules and embedding datasets. The project describes itself as alpha and meant for research, so it is best treated as one layer in a defense-in-depth setup rather than a complete fix.

As a guardrail framework, Vigil sits between your application and the model. It combines several techniques in one pipeline: vector similarity against known attacks, YARA heuristics, a transformer classifier, prompt-response similarity, canary tokens, and sentiment analysis. Detections from each scanner are reported together so you can decide how to handle a flagged prompt.

// What it does

• Modular, extensible scanners you can mix and match per pipeline
• YARA-based heuristic signatures, plus support for your own custom rules
• Vector database / text similarity scanning, with optional auto-updating on detected prompts
• Transformer model classifier and prompt-response similarity checks
• Canary tokens and sentiment analysis scanners
• Runs as a Python library or a REST API server, with a Streamlit web UI playground

// Getting started

Vigil installs from source. Clone the repo, install YARA, set up a virtualenv, install the library, then use it from Python or run the API server.

git clone https://github.com/deadbits/vigil-llm.git
cd vigil-llm

python3 -m venv venv
source venv/bin/activate
pip install -e .

from vigil.vigil import Vigil

app = Vigil.from_config('conf/openai.conf')

result = app.input_scanner.perform_scan(
    input_prompt="prompt goes here"
)

python vigil-server.py --conf conf/server.conf

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Screen user prompts for known prompt-injection and jailbreak patterns before they reach your LLM
• Add a guardrail microservice in front of an existing chatbot or RAG app via the REST API
• Maintain custom YARA signatures and embedding datasets for attacks specific to your product
• Use canary tokens and prompt-response similarity to catch data exfiltration or leaked system prompts

// How Vigil compares

Vigil alongside other open-source guardrails & security tools AI/TLDR tracks, ranked by GitHub stars.