ARES Usage

Note

ARES requires Python 3.10+.

Setup

git clone https://github.com/IBM/ares.git
cd ares
python -m venv .venv
source .venv/bin/activate
pip install .

Note

Optional extras: use .[dev] for development (tests, linting) or ares-redteamer[dev] for a pinned version.

Create a config file my-example.yaml:

target:
  huggingface:

red-teaming:
  prompts: assets/pii-seeds.csv

Run ARES:

ares evaluate example_configs/my-example.yaml --limit

To learn how to configure ARES for different use cases, see the ARES Configuration.

Limiting Attack Goals

To limit the number of attack goals tested, use the --limit and --first options:

ares evaluate example_configs/minimal.yaml --limit         # Limits to first 5 goals
ares evaluate example_configs/minimal.yaml --limit --first 3  # Limits to first 3 goals

Dashboard Visualization

ARES includes a built-in dashboard to visualize configurations and evaluation results.

To enable the dashboard during evaluation:

ares evaluate example_configs/minimal.yaml --dashboard

To visualize a report independently after evaluation:

ares show-report example_configs/minimal.yaml --dashboard

Notebook

ARES can also be run interactively via Jupyter notebooks. See:

notebooks/Red Teaming with ARES.ipynb

This notebook walks through the Goal, Strategy, and Evaluation components in detail.

Assets

ARES uses datasets stored in the assets/ folder. To run example configs, download the following file from HarmBench:

harmbench_behaviors_text_all.csv from: https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_all.csv

Place it in the assets/ directory.

Customizing the Target Model

You can change the target model by editing the config file. For example:

target:
  huggingface:
    model_config:
      pretrained_model_name_or_path: ibm-granite/granite-3.3-8b-instruct
    tokenizer_config:
      pretrained_model_name_or_path: ibm-granite/granite-3.3-8b-instruct

Refer to example_configs/connectors.yaml for more options.

Plugins

ARES supports a modular plugin architecture that allows you to extend its functionality without modifying the core codebase.

Plugins can be used to:

Connect to external targets (e.g., APIs, LLMs, agentic apps)
Define custom attack goals
Implement new attack strategies
Evaluate model responses with specialized metrics

To learn how to install, configure, and create plugins, see ARES Plugins.

Next Steps

Explore example configs in example_configs/
Try different goals, strategies, and evaluators
Visualize results using the built-in dashboard

Common Issues

Missing `.env` file

Create one with required keys (e.g., API tokens).
`ModuleNotFoundError`

Run:
```
pip install .
pip install .[dev]
```
Dashboard not launching
Check for:
Port conflicts

Missing dependencies

Plugins not available

Ensure plugins are installed, for example:
pip install plugins/ares-garak
Or use ARES built-in installer:
ares install-plugin ares-garak
ares install-plugin ares-human-jailbreak