Red teaming

Write the attacks your AI system will face

Adversarial prompts and test cases for your own assistant or agent, each with expected behaviour and a fail signal. Unrestricted models write the attack side typical assistants decline, within the Acceptable Use Policy.

Get API key Read the quickstart

Unrestricted means fewer refusals, not no rules. Read the Acceptable Use Policy

Test material your harness can run.

Ask for cases in a fixed shape and they go straight into your suite.

  • cases/PI-0142.yaml Test case

    A case with an expected outcome

    Input, expected behavior and a failure condition your scorer can check.

  • /v1/chat/completions JSON mode

    Variants from one seed

    Turn a seed case into variants across tone, language and wrapping, returned as JSON your runner loops over.

  • coverage.md Report

    Coverage notes

    Paste your run results and get a summary of what the suite covers and where it is thin.

From system prompt to regression suite.

The model writes the cases. Sending them to your system and scoring the replies happens in your own code.

Read the quickstart

  1. You

    Describe the system under test

    Its purpose, its tools and the policy it must follow.

  2. Modellane

    Writes cases in your schema

    Input, expected behaviour and fail signal per case. JSON mode keeps them parseable.

  3. You

    Run them in your harness

    Send each case to your system and score the replies with your own checks.

  4. Modellane

    Writes variants around failures

    Tone, language and wrapping change; the fail signal stays the same.

  5. You

    Keep what fails

    Failing cases join your regression suite before the next release.

What we do, and what we don't.

Modellane returns text. Everything that touches your system stays on your side.

Read the Acceptable Use Policy

What we do

  • Serve models that write adversarial prompts and test cases
  • Return them as plain text or JSON in the shape you ask for
  • Bill per token from a prepaid balance

What we don't do

  • Run your harness, score results or keep your suites
  • Send traffic to your system or anyone else's
  • Store your prompts or the model's replies
  • Allow uses the Acceptable Use Policy prohibits, whatever the model answers

Which models fit: Models whose id contains -unrestricted write adversarial prompts that typical assistants refuse or water down. Standard models keep their base model's behavior, behind the same API. See every model

FAQ

Red teaming questions

Do you run the tests against my system?
No. The API returns text. Your harness sends the cases to your system and scores the replies.
Which systems can I test?
Systems you own or are authorized to test. The Acceptable Use Policy applies to every request, including test generation.
Why not use a typical assistant to write test cases?
Typical assistants often decline to write the attack side of a test, or soften it until it no longer tests anything. Unrestricted models write it as asked.
Are my test cases stored?
No. Prompts and replies are never stored. We keep a record without content for 180 days for billing.

Test before your users do.

Create a key, describe your system and generate the first suite.

Get API key Read the quickstart