Installation¶

pikit requires Python 3.9+ and has a zero-dependency core — attacks, defenses, channels, and the mock target work out of the box with no third-party packages.

From source (recommended)¶

git clone https://github.com/NY1024/pikit.git
cd pikit

pip install -e .                 # core: attacks + defenses + channels (zero deps)

Optional extras¶

Model SDKs are installed only when you need a specific backend:

pip install -e ".[openai]"       # + OpenAI / OpenAI-compatible (vLLM, Ollama, DashScope)
pip install -e ".[anthropic]"    # + Anthropic Claude
pip install -e ".[hf]"           # + local HuggingFace transformers
pip install -e ".[all,dev]"      # everything + pytest

Extra	Packages	When to install
`openai`	`openai>=1.0`	OpenAI API, vLLM, Ollama, DashScope/Qwen, LM Studio
`anthropic`	`anthropic>=0.39`	Anthropic Claude models
`hf`	`torch>=2.0`, `transformers>=4.40`	Local HuggingFace models
`dev`	`pytest>=7.0`	Running the test suite

Verify the install¶

python -c "from pikit import attacks, defenses, channels; print(attacks.list())"
# ['combined', 'context_ignoring', 'escape', 'fake_completion', 'naive', 'obfuscation', 'payload_splitting', 'prefix_injection', 'prompt_leaking']

Run the test suite¶

The full test suite is offline (no API key required):

pytest

Configuring model access¶

Real backends read credentials from the environment — never hardcode a key or pass it on the command line.

cp .env.example .env            # then edit .env with your real key
set -a; source .env; set +a     # export the vars into your shell

Variable	Meaning
`OPENAI_API_KEY`	Key for an OpenAI-compatible endpoint
`OPENAI_BASE_URL`	Endpoint URL; omit for OpenAI. DashScope/Qwen: `https://dashscope.aliyuncs.com/compatible-mode/v1`
`ANTHROPIC_API_KEY`	Key for Anthropic Claude
`PIKIT_MODEL`	Default model id for the demos (optional)

.env is gitignored. If a key was ever pasted on a command line or committed, rotate it — a leaked secret can't be un-leaked.