Getting Started

PentestLoop uses evolutionary red-teaming to find weaknesses in AI agents. Here's how to run your first adversarial campaign in three steps.

1

Register Your Agent

Connect your AI agent so PentestLoop can test it. Supports HTTP APIs, CreatorsAGI Envoy, or use the built-in mock agent to try it out without a real target.

  • •HTTP API — Any AI agent with a REST endpoint. Send a message, get a response.
  • •Mock Agent — Built-in simulated agent for exploring PentestLoop risk-free.
  • •Envoy — CreatorsAGI AI Envoy integration for testing deployed companions.
Register Agent
2

Configure a Campaign

Set up an adversarial test run. Choose what to test (security, quality, compliance), how large the attack population should be, and how many generations to evolve.

  • •Population Size — How many attack strategies compete each generation.
  • •Max Generations — How many evolution cycles to run.
  • •Test Dimensions — Security (injection, leakage), Quality (hallucination, reasoning), Compliance (policy, bias).
Create Campaign
3

Launch & Monitor

Watch as PentestLoop evolves attacks in real time. The fitness chart shows how effective attacks become over generations, and vulnerabilities are flagged as they're discovered.

  • •Fitness chart tracks attack effectiveness over generations.
  • •Generation browser lets you inspect individual attacks and conversations.
  • •Vulnerabilities are categorized by severity — critical, high, medium, low.
View Dashboard
▶

Try It Now — Run a Demo Campaign

Don't have an agent to test yet? Run a demo campaign against our built-in mock agent. It simulates a sales AI with configurable vulnerabilities — no API keys or real agents needed. You'll see real evolutionary red-teaming with conversations, scoring, and vulnerability detection.

Key Concepts

Generation—One evolution cycle — attacks are tested, scored, and the best ones mutate to create the next generation.
Fitness—A 0–1 score measuring attack effectiveness. Higher = more successful at exposing weaknesses.
Vulnerability—A confirmed weakness — data leakage, jailbreak, safety bypass, or policy violation.
Population—The pool of attack strategies competing each generation. Larger populations explore more strategies.
Seed Attack—A starting attack strategy. Seeds are evolved over generations into more sophisticated attacks.
Convergence—When attacks stop improving significantly. The campaign ends early if fitness plateaus.