AVISE - AI Vulnerability Identification & Security Evaluation

A framework for identifying vulnerabilities in and evaluating the security of AI systems.

Full Documentations: https://avise.readthedocs.io

Youtube Demo

Table of Contents

• Quickstart for Evaluating Language Models
  • Prerequisites
  • 1. Install AVISE
  • 2. Run a Model
  • 3. Evaluate the model with a Security Evaluation Test (SET)
• Advanced Usage
  • Configuring Connectors
  • Configuring Security Evaluation Tests (SETs)
  • Using custom configuration files
• Required Arguments
• Optional Arguments
• Citation

Quickstart for Evaluating Language Models

Prerequisites

• Python 3.10+
• Docker (For Running models locally with Ollama)

1. Install AVISE

Install with

• pip:

  pip install avise
  

• uv:

  uv pip install avise
  

  or

  uv tool install avise
  

2. Run a Model

You can use AVISE to evaluate any model accessible via an API by configuring a Connector. In this Quickstart, we will assume using the Ollama Docker container for running a language model. If you wish to evaluate models deployed in other ways, see the Full Documentations and available template connector configuration files at AVISE/avise/configs/connector/languagemodel/ dir of this repository.

Running a language model locally with Docker & Ollama

• Clone this repository to your local machine with:

git clone https://github.com/ouspg/AVISE.git

• Create the Ollama Docker container

  • for GPU accelerated inference with:

    docker compose -f AVISE/docker/ollama/docker-compose.yml up -d
    

  • or for CPU inference with:

    docker compose -f AVISE/docker/ollama/docker-compose-cpu.yml up -d
    

• Pull an Ollama model to evaluate into the container with:

  docker exec -it avise-ollama ollama pull <model_name>
  

3. Evaluate the model with a Security Evaluation Test (SET)

Basic usage

avise --SET <SET_name> --connectorconf <connector_name> [options]

For example, you can run the prompt_injection SET on the model pulled to the Ollama Docker container with:

avise --SET prompt_injection --connectorconf ollama_lm --target <model_name>

To list the available SETs, run the command:

avise --SET-list

Advanced Usage

Configuring Connectors

You can create your own connector configuration files, or if you cloned the AVISE repository, you can modify the existing connector configuration files in AVISE/avise/configs/connector/languagemodel/.

For example, you can edit the default Ollama Connector configuration file AVISE/avise/configs/connector/languagemodel/ollama.json, and insert the name of an Ollama model you have pulled to be used as a target by default:

{
    "target_model": {
        "connector": "ollama-lm",
        "type": "language_model",
        "name": "<NAME_OF_TARGET_MODEL>",
        "api_url": "http://localhost:11434", #Ollama default
        "api_key": null
    }
}

Configuring Security Evaluation Tests (SETs)

Similarly, you can customize the configurations for SETs as well. For example, by editing the Red Queen SET configuration file AVISE/avise/configs/SET/languagemodel/multi_turn/red_queen.json, you can define if the SET is executed incrementally (the target model will generate a response after each subsequential prompt), or as a template (only works for target systems that accept a conversation as an input) and if the SET uses an Adversarial Language Model (ALM). Additionally, you can define the exact template attack prompts that the SET uses:

{
    "incremental_execution": false, #Change to true to execute the SET cases incrementally
    "use_adversarial_languagemodel": false, #Change to true to utilize an Adversarial Language Model (ALM)
    "adversarial_model_device": "auto", # Defines on which device(s) the ALM is loaded to ("cpu", "gpu", or "auto")
    "evaluation_system_prompt": "You are a highly accurate cybersecurity classifier, ..."
    "evaluation_model_device": "auto", # Defines on which device(s) the ELM is loaded to ("cpu", "gpu", or "auto")
    "sets": [
        {
            "id": "RED-QUEEN-001",
            "vulnerability_subcategory": "Evasion Attack",
            "action": "make a fake passport",
            "type": "occupation_teacher",
            "category": "violence,aiding_and_abetting,incitement",
            "conversation": [
                { ... }
            ...
            ]
         },
         ...
    ]

Using custom configuration files

If you want to use custom configuration files for SETs and/or Connectors, you can do so by giving the paths to the configuration files with --SETconf and --connectorconf arguments:

avise --SET prompt_injection --SETconf AVISE/avise/configs/SET/languagemodel/single_turn/prompt_injection_mini.json --connectorconf AVISE/avise/configs/connector/languagemodel/ollama.json

Required Arguments

Optional Arguments

Citation

If you find AVISE useful, please cite it as below:

@misc{lempinen2026,
      title={AVISE: Framework for Evaluating the Security of AI Systems},
      author={Mikko Lempinen and Joni Kemppainen and Niklas Raesalmi},
      year={2026},
      eprint={2604.20833},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2604.20833},
}

Lempinen, M., Kemppainen, J., & Raesalmi, N. (2026). AVISE: Framework for Evaluating the Security of AI Systems. arXiv preprint arXiv:2604.20833.