Run a first security investigation

Build a small evidence set, choose a mode and turn a broad question into a checkable next step.

A productive first conversation starts with a question small enough to answer from the material at hand. “Review our security” hides the system, threat and expected result. “Explain why this report handler accepts a non-owner” gives the investigation a boundary.

Keep the original evidence intact. A model’s summary helps you navigate it, but the source remains the place to verify a claim.

Before you start

  • One question, the affected component and the expected behavior.
  • A selected text file, source excerpt or sanitized request/response pair. Remove credentials and unrelated personal data.
  • An account with current access and enough available budget for the request. Check model and mode availability in the application.

Follow the steps

  1. Write the invariant you expect the system to enforce and the observation that appears to contradict it.
  2. Choose Chat for discussion, Code for selected source or the available specialized mode for its documented task. Model selection and mode selection are separate decisions.
  3. Attach only the material needed for the first question. Ask the model to label direct observations, interpretations and missing evidence.
  4. Follow one unresolved question at a time. When an action is proposed, inspect its exact inputs, scope and budget before approval.
  5. Record the result and what would disprove it. Expand context only when the next step needs it.

Use a precise prompt

Adapt to your evidence
Our expected rule is that only a report’s owner can read its body. Review the attached handler and two sanitized responses. Summarize direct observations first, then possible explanations. Identify the smallest next check that distinguishes those explanations and state what result would disprove the suspected flaw.

Verify your work

  • Each conclusion points to a supplied source or a recorded tool result.
  • Missing files, unavailable tools and untested assumptions remain visible.
  • You can state the next action, its expected result and its stopping condition before running it.

Bring your own evidence into a conversation and work through the next question.

Open chat