Red-team your AI agent. No security team required.

Sign upRead the docs
Open source engines
3+
Attack categories
50+
Used by
Builders
  • INJ
  • JBK
  • LEK
  • TLM
  • EXF
  • PRV
BRD Your agent
ENDEndpointMDLModelSYSSystem promptTLSToolsMEMMemoryJBKINJTLMPRVLEKEXF
Each channel reaches a different part of your agent.
  • MDL

    Model

    The language model your agent reasons with, and the one attackers try to talk out of its instructions.

  • SYS

    System prompt

    The instructions that define what your agent is allowed to do. The first thing an attack tries to read or rewrite.

  • TLS

    Tools

    Every function, API, or shell your agent can call. Each one is a way to make it act, not just speak.

  • MEM

    Memory

    What your agent remembers across turns and sessions, and what an attacker can plant there or pull out.

  • END

    Endpoint

    The HTTP surface your agent answers on, with its own auth and its own request shape. This is where every attack enters.

  • INJPrompt injection
  • JBKJailbreak
  • LEKData leakage
  • TLMTool misuse
  • EXFExfiltration
  • PRVPrivilege escalation

What gets attacked

An agent is more than a model. It has instructions, tools it can call, memory it keeps, and an endpoint it answers on. Each one is a place an attack can enter or a thing an attack can take.

You say what you want to break. The platform picks the channels that reach it and runs them against your agent, without you learning a single tool first.

Three steps, no setup beyond your keys

Point it at your agent's endpoint, bring your own model key, and it runs real attacks from established red-teaming engines for you.

The workflow stays the same as more engines and options arrive. Define what to hit, attack it, read what happened.

  1. 01

    Define target

    Give it your agent's endpoint and auth key, plus your own model API key. That is the whole setup.

  2. 02

    Attack

    It runs attacks across the channels above, using established red-teaming engines, on your behalf.

  3. 03

    Report

    See which attacks got through and where, with the exchanges that show it.

Real engines, run for you

There is no secret attack technique here. The platform orchestrates established open-source red-teaming engines, points them at your target, and handles the adapter work between your endpoint and each engine so you never have to.

  • Available

    garak

    NVIDIA's open-source LLM vulnerability scanner.

  • Available

    promptfoo

    Open-source red-team and evaluation toolkit.

  • Planned

    PyRIT

    Microsoft's open-source red-team toolkit.

Point it at your agent. Find out what breaks before someone else does.

Sign upRead the docs