redteam-cli 0.1.0

Local-first red-team scanner for ML models — runs adversarial, extraction, and prompt-injection attacks on your machine and uploads only the scores.

redteam-cli

Local-first security scanner for ML models. Runs five adversarial / extraction / prompt-injection attacks against your model on the machine that already has it and uploads only the per-attack scores. The model itself never leaves your environment.

Install

pip install "redteam-cli[torch]"     # image classifiers (.pt / .pth / state-dict)
pip install "redteam-cli[onnx]"      # ONNX models (boundary + extraction)
pip install "redteam-cli[hf]"        # HuggingFace LLMs via hf:org/model
pip install redteam-cli              # CLI only (use --dry-run, no attacks)

Usage

export REDTEAM_API_KEY=rtai_…
export REDTEAM_API_URL=https://your-redteam-platform.example.com

# Image classifier (full nn.Module saved with torch.save)
redteam scan --model ./mymodel.pt --type image-classifier \
    --input-shape 3,224,224 --num-samples 32

# State-dict only — supply the architecture file
redteam scan --model ./weights.pt --type image-classifier \
    --arch ./model.py:MyResNet --input-shape 3,224,224

# ONNX — boundary + extraction (FGSM N/A, no gradients)
redteam scan --model ./mymodel.onnx --type image-classifier --input-shape 3,224,224

# LLM via HuggingFace
redteam scan --model hf:Qwen/Qwen2-7B-Instruct --type llm

# LLM via custom .py adapter
redteam scan --model ./llm_adapter.py --type llm

# Dry-run anything (run locally, print JSON, skip upload)
redteam scan --model ./mymodel.pt --type image-classifier --dry-run

Attacks

Attack	Model types	What it tests
FGSM	image-classifier	White-box adversarial robustness (single ε-step gradient sign)
Boundary	image-classifier	Black-box decision-boundary fragility (median L2 to flip)
Model extraction	image-classifier, tabular	Whether a surrogate trained from queries can clone the model
Prompt injection	llm	Canary-leak detection across role-hijack / jailbreak payloads
Output manipulation	llm	Refusal-marker matching on disallowed prompts (markers customisable for non-English models)

Where the results go

This CLI sends results to the redteam-ai platform (FastAPI backend + Next.js dashboard) running either locally via Docker Compose or deployed to your own host. To run the platform yourself:

git clone https://github.com/Anbar26/redteam-ai
cd redteam-ai && docker compose up -d --build
# Visit http://localhost:3000 — sign up to get an API key

Or use --dry-run to skip the upload entirely and just read the JSON.

License

MIT