Set up and use Security Threat Model safely

This guide explains how teams can prepare the official OpenAI Security Threat Model skill in Codex and use it for a controlled security analysis. It follows the primary SKILL.md and assumes that the skill is obtained from the official repository.

Define Scope Before Analysis

Start by identifying the exact repository root or project path that is in scope. Record its purpose, deployment model, expected authentication, possible internet exposure, and known external services. Also document directories that are explicitly out of scope. A narrow boundary improves traceability and prevents test fixtures or local helper programs from being treated as production architecture without evidence.

Collect Evidence Safely

Ask the agent to derive components, data flows, and entry points only from files that are actually present. Useful evidence includes paths, symbols, configuration keys, and short non-sensitive excerpts. Do not expose secret configuration values in the report. If tokens, passwords, or keys are encountered, they must be redacted; the report may describe only that a secret exists and where it was found. Before analysis, confirm that the working copy, logs, and generated report are held in an appropriately protected environment.

Analyze and Resolve Questions

The skill organizes trust boundaries, assets, realistic attackers, and abuse paths. Require each threat to include likelihood, impact, priority, and a short rationale. Then review the one to three targeted questions about operations, authentication, data sensitivity, or tenant separation. Confirmed answers should become part of the report context. If answers remain unavailable, retain the assumptions and explain how they influence prioritization.

Review and Share the Report

Check that runtime, CI, build, development, tests, and examples remain distinct. Every discovered boundary and entry point should appear in at least one threat or have a documented reason for being irrelevant. Ensure the Mermaid diagram stays compact and that existing controls are separated from recommended mitigations. Save the result using the repository or directory name with the suffix threat-model.md, then ask a responsible security professional to review it before distribution.

Workflow Security Boundaries

The skill produces a strong working baseline, but it does not automatically certify a system or replace independent testing. Recommendations must be checked against actual operations, access controls, network paths, and data classification. Avoid uncontrolled changes to the target system and perform validation in an isolated environment. A shared report must not contain secrets, personal data, or architectural claims that have not been supported by repository evidence.