respan-redteam 0.1.0

Adaptive red-team engine and CLI for AI agents

Respan Red Team

Open-source red-team engine and CLI for AI agents. It probes an agent through normal chat messages, adapts its attacks, judges the responses, and produces an OWASP-aligned report.

Only scan systems you own or have permission to test.

Install

Requires Python 3.11 or newer.

pip install respan-redteam

For development:

git clone https://github.com/respanai/respan-redteam.git
cd respan-redteam
uv sync

Connect your agent

Create adapter.py with a TARGET. A target opens conversations, and a chat sends user messages to your agent:

class Chat:
    def send(self, message: str) -> str:
        return my_agent(message)

    def transcript(self) -> list[dict]:
        return []


class Target:
    label = "my-agent"

    def open(self) -> Chat:
        return Chat()


TARGET = Target()

Complete templates are available for a local adapter and a server-owned session.

Run

The default remote scan uses Respan's hosted engine:

respan-redteam adapter.py -o report.json

To run the engine locally, provide an OpenAI-compatible API key:

export OPENAI_API_KEY=...
respan-redteam adapter.py --local -o report.json

From a source checkout, use uv run respan-redteam. Run respan-redteam --help for output, retry, CI-grade, and connection options.

Python API

from respan_redteam import run_campaign
from adapter import TARGET

result = run_campaign(TARGET)
print(result.grade(), result.score())
print(result.to_report())

Extend

Custom prompt attacks, payload carriers, and multi-step strategies can be registered through the public extension API. See the prompts, carriers, and strategies packages.

Local model settings use OPENAI_API_KEY, optional OPENAI_BASE_URL, and the RESPAN_MODEL_* variables documented in config.py.

Development

just test

Licensed under Apache 2.0.