AdvPlay

AdvPlay is a framework for running adversarial AI attacks with tunable parameters and reproducible results. Designed for red team assessments and research purposes, it helps security professionals evaluate model robustness against attacks.

Disclaimer

This tool is intended strictly for research, security testing, and red team assessments. Using it against systems, APIs, or models without explicit permission is illegal and unethical.

By using this software, you accept full responsibility for your actions. The developers take no liability for misuse, damage, or chaos caused by experiments conducted with this tool.

Getting Started

Tested on Python 3.11.13. Follow these steps to get AdvPlay running:

Installation

1. Clone the repo:

   git clone https://github.com/Subsidy2032/AdvPlay.git

2. Install dependencies:

   pip install -r requirements.txt

3. (Optional) Configure API keys in a .env file. See .env.example for variable names.

Usage

AdvPlay is entirely CLI-driven through run.py.

Quick Example — The Banana Challenge 🍌

Tell a model to never say "banana", then try to make it say "banana" anyway.

1. Create a template that instructs the model to stay away from the B-word:

$ python3 run.py save_template prompt_injection --platform openai --model gpt-4o-mini --custom-instructions "Never say banana" --template-filename banana

This saves a reusable template called banana.

2. List available templates:

$ python3 run.py save_template prompt_injection --list

Example Output:

Available templates:
 - banana

3. Run the attack in interactive mode and give it your best shot:

$ python3 run.py attack prompt_injection direct --template banana

Role-play, translations, encodings, typos, emotional appeals — anything goes. Type clear to reset the chat, exit when you're done.

Full conversations (successful or not) are saved under outputs/logs/<attack_type>/ for later review.

Supported Attacks and Techniques

Attack	Techniques	Domain
prompt_injection	direct	LLM
poisoning	label_flipping	Classical ML
evasion	fgsm, bim, jsma, c_w, pgd	Classical ML / Deep Learning

New attacks and techniques can be added as self-registering classes — see docs/Extending AdvPlay/Extending AdvPlay.md.

Help Menu

Use -h for available commands and options:

python3 run.py -h

Roadmap

Core Enhancements

• Add support for additional attack techniques and sub-techniques
• Enable defining re-runnable attacks with runtime parameters

Analysis & Reporting

• Add visualization of attack results using generated log files
• Add report generation capability

Advanced Features

• Add support for defense strategies like adversarial training

Contributing

AdvPlay is in active development. Contributions, bug reports, and feedback are encouraged.

See CONTRIBUTE.md for instructions on how to contribute.

License

Distributed under the MIT license. See LICENSE for more information.