Security
Treat a value as trusted only because of who wrote it. Write the threat model and the abuse-case tests before the feature. Leave the result in chat, or as a section of the design doc. Do not create a security-output folder.
Quick start
- Pick a mode: building across a trust boundary, triaging a finding, or reviewing a design.
- Stop for a yes on anything in the ask-first list.
- Run the checks the project already has, plus the two checks under Scanners.
- Hand over what ran, what did not, and a "Could improve" list.
Building across a trust boundary
Use this when the change accepts input, decides who may act, sets a session, takes an upload, fetches a URL, reads a secret, opens a file named by data, or handles model output.
- In chat, name the trust boundaries, the assets, the STRIDE letters for each interaction, and 2–5 abuse cases. Trust follows who wrote a value, not which channel delivered it. Read threat-model.md.
- Turn each abuse case into a test and watch it fail, then write the code. Why: a control you have not seen fail is unproven.
- Apply the controls in controls.md. A model feature also needs llm.md.
- Personal, payment, or health data also needs privacy.md.
Example. "Fetch the title of a URL the user pastes." Caller to process: all of STRIDE. Process to outbound call: tampering, information disclosure, denial of service, and the caller picks the destination. One abuse case: the URL resolves to a link-local address. The test expects that fetch to be refused, and it fails before the check exists.
Triage a finding
A scanner hit, an advisory, or a reported bug. For each one, say whether the code is on a runtime, build, test, or deploy path, whether a fix is published, and which cell of the impact × likelihood grid it sits in. Propose one fix and apply it after a yes. A deferral records a reason and a review date. Do not run a forced remediation (npm audit fix --force or the equivalent). Why: it can change dependency ranges the project did not choose. Steps: dependencies-and-scanning.md.
Design review
Write the STRIDE table, tag each control preventive, detective, or corrective, and write the residual risk. Add that section to the design doc through the design-docs skill, using the template in threat-model.md. Why: a review that exists only in chat disappears.
Ask first
Stop for a yes before a new or changed auth flow, a new category of personal or sensitive data, a CORS change, an upload handler, a rate-limit change, granting a role or permission, or a new external integration. Why: each of these is hard to take back once it is live.
Scanners
Run the scanners the project already configured. On a security-relevant change, also run the ecosystem's known-vulnerability check when that command is already available (govulncheck as the go skill describes, npm audit, pnpm audit, pip-audit, or cargo audit), and this secrets check, which prints file names only:
git diff --cached -i -G'password|secret|api_key|token' --name-onlyA printed name is a file to open before you commit. Why: the check must not copy a secret into the tool log. Do not install a scanner, add one to CI, add a config, or suppress a finding unless asked. A suppression names the rule and the reason. Commands and the optional scanner set: dependencies-and-scanning.md.
Hand over
- What ran, and the result. What you could not run, and why.
- The threat model, the triage, or the design-doc section.
- Could improve: each gap with a one-line setup. Propose it, and apply it after a yes.
Replies
Talk to the user in simple English: short sentences, everyday words, and a technical term only where it is needed, explained the first time it appears.
References
| File | Read when |
|---|---|
| threat-model.md | starting a feature, rating a finding, or writing the design-review section |
| controls.md | implementing a control |
| llm.md | the feature calls a model or puts untrusted text in a prompt |
| dependencies-and-scanning.md | a finding, a new dependency, a release, no scanners, or choosing the audit command |
| privacy.md | personal, payment, or health data |