Community / AI red teaming

The field guide.

A starting point for curious minds. Learn the language, find a place to practise, and make your findings useful to someone else.

An independent resource from NestCipher. Gray Swan and the other projects below are separate organisations.

01 / A research workflow

Make every attempt
teach you something.

Start with a question. Keep the evidence. Change one variable at a time.

  1. Understand the failure mode.

    Pick a risk and learn what success or failure looks like. Our OWASP explorer connects the definitions with examples and mitigations.

    Open the OWASP explorer

  2. Choose a scoped challenge.

    Find a suitable Arena challenge, read its rules, and note the target behaviour. Establish a baseline before testing your hypothesis.

    Explore Gray Swan Arena

  3. Document the observation.

    Record the exact input, output and test conditions. Repeat the attempt and distinguish a reproducible failure from a one-off response.

    Open the Research Workbench Use the notes checklist

Repeat the loop ↻ Keep the scope, change one variable, and compare with your baseline.

02 / Practise with known evidence

AUTHORED EXERCISES / VERSION 1

Read the record.
Question the conclusion.

Work through invented cases with inspectable evidence and written explanations. Then prepare an untested, private experiment in the Workbench. No model or tools run in these labs.

01

Who can authorise this?

Separate a source's claim from permission to act.

02

What actually happened?

Look beyond the final reply to assess the observed effect.

03

Is this a useful comparison?

Separate changed conditions from the conditions you do not know.

03 / Keep a useful record

A finding is more
than a screenshot.

A simple record makes an experiment easier to reproduce, compare and explain.

RESEARCH NOTESPLAIN-TEXT CHECKLIST
target
Model, version, date and permitted scope
hypothesis
The specific behaviour you are testing
input
Exact prompt and relevant context
observation
Actual response, proposed actions and completed effects
reproduction
Attempts, conditions and repeatability
impact
Why the observed failure matters

04 / Outside the nest

Good places
to go deeper.

EXTERNAL RESOURCES ↗
PRACTISE

Gray Swan Arena

app.grayswan.ai

Explore structured AI red-teaming challenges and the community around them. Read each challenge’s rules and scope before you begin.

REFERENCE

OWASP GenAI Security

genai.owasp.org

The primary reference for the LLM Top 10, including prompt injection, sensitive information disclosure and excessive agency.

INVESTIGATE

PortSwigger Research

portswigger.net

Follow technical security research and see how researchers explain a vulnerability, its impact and a reproducible technique.

UTILITIES

CyberChef

gchq.github.io

Decode, transform and inspect data with recipes you can save and share. Useful when an input or output needs a closer look.

These resources open on their own websites. Challenge availability and access are managed by their respective providers.

05 / Build the field guide

PUBLIC TEACHING MATERIAL

Teach a lesson.
Show your reasoning.

A useful contribution starts with an invented example, a precise question and evidence someone else can inspect. Use the template to prepare a teaching proposal for maintainer review.

  1. 01

    Invent the scenario. Use made-up tasks, records and recipients. Leave private prompts, transcripts and challenge techniques out.

  2. 02

    Expose the evidence. Distinguish a reply, a proposed action and a completed effect. State what remains unknown.

  3. 03

    Explain the answer. Include a counterexample, the limits of the conclusion and a primary reference.

Download the lab proposal template

A Markdown template to work on locally. This site does not accept or publish submissions. These lessons are reviewed as authored teaching examples, not measurements of a model's behaviour.

Something missing from your research workflow?

Suggest a tool or resource