| name | secure-design |
| description | Threat-model a code, configuration, infrastructure, auth, data, cryptography, networking, or dependency change by decomposing it into application sub-components and applying each one's threat lens and controls. Use when building or modifying a feature with security in scope, or when the global Security Gate is accepted. Not for incident investigation (use rca-investigation) or for writing security tests alone (use adversarial-test-design). |
| version | 1.1.0 |
Secure Design
Use this skill when modifying code, configuration, infrastructure, or dependencies and security is in scope. Assume hostile input, compromised dependencies, misconfigured infrastructure, and least privilege by default.
Philosophy: decompose by sub-component
Decompose the change into the application sub-components it touches, then apply each sub-component's threat lens and controls. The sub-component is the unit of analysis: it maps to where code and config actually live, bounds the review, and routes each part to a specific, concrete threat profile.
Sub-components:
- Business Logic
- Authentication and Authorization
- Data, Secrets & Privacy
- Cryptography
- Supply Chain
- Infrastructure
- Networking
- Observability and Logging
Method
- Identify the change surface: which files, services, endpoints, infrastructure, or dependencies change.
- Map it to the affected sub-components above. Most changes touch two or three; ignore the rest, but say which you ignored.
- For each affected sub-component:
- Identify the trust boundaries it crosses (untrusted → trusted) and the sensitive data that flows through it.
- Apply that sub-component's threat lens and controls below.
- Name the security assumptions you depend on but did not verify.
- Run the pre-PR threat-model checklist for the touched sub-components.
- Add or update security tests for the risky paths.
- Run security tooling when available and allowed, before pushing upstream.
- If unsure about a technology use web-research to find details and audit them before blindly trusting content.
Rule: Data that crosses a trust boundary is untrusted until explicitly validated at that boundary, not deep in business logic. Validation, errors, logs, and responses at a boundary must not leak sensitive data back across it.
1. Business Logic
Threats: injection, abuse of legitimate flows, state/workflow corruption, validation buried below the boundary.
- Validate every input surface (HTTP, CLI, environment variables, files, IPC, queues, webhooks, and database values reused in queries): check type, length, format, range, and encoding before use. Prefer allowlists over denylists. Normalize before validating when equivalent encodings may exist.
- Never concatenate untrusted data into SQL, shell, HTML, templates, LDAP, or XPath. Parameterize queries. Treat file paths as untrusted: resolve and jail them to an expected root before access.
- Check that multi-step flows cannot be reordered, replayed, or partially completed to reach a state the design forbids (skipping payment, re-using a one-time action, downgrading a check).
Red flags: string interpolation into SQL, shell, HTML, templates, LDAP, or XPath; path traversal via filenames or archive extraction; unsafe deserialization; user-controlled regex patterns; user-controlled template rendering.
2. Authentication and Authorization
Threats: missing or bypassable authorization, IDOR, privilege escalation, session fixation/hijacking, trusting client-supplied identity.
- Treat authentication (who is the caller) and authorization (what the caller may do) as separate checks. Default to deny; grant explicitly.
- Enforce authorization close to the protected resource, not only at the routing or controller layer.
- Derive identity and tenancy from server-side session or verified token claims, never from client-supplied IDs, headers, or request bodies.
- Re-check authorization on every object access, not just the listing or entry point (IDOR). Cover cross-user, cross-tenant, and cross-org access.
- Session tokens must be long, random, and CSPRNG-generated, and invalidated server-side on logout, credential rotation, or privilege change.
Rule: A function that touches a protected resource must either enforce permissions itself or document the upstream permission invariant it depends on.
3. Data, Secrets & Privacy
Threats: secret leakage, over-collection, sensitive data in logs, unencrypted sensitive data.
- Never hardcode secrets, API keys, passwords, tokens, or private/signing keys in source or config. Use environment variables or a secret store; reference secrets by name only.
- Never print secrets or unnecessary PII in logs, errors, traces, telemetry, metrics, or test snapshots. Mask sensitive values before logging.
- Exclude secret-bearing files via
.gitignore, .dockerignore, .claudeignore, or equivalent.
- Identify whether the change handles PII, credentials, tokens, or financial, health, customer, or internal security data. Minimize collection and retention. Encrypt sensitive data at rest and in transit. Use synthetic or anonymized test data; do not copy production data into dev or test.
On discovery: if you find a hardcoded secret, stop and flag it immediately. Do not silently move, reuse, or normalize it. Recommend rotation if exposure is plausible.
Rule: Sensitive data needs a clear owner, purpose, retention period, and access boundary.
4. Cryptography
Threats: weak or broken primitives, key and nonce misuse, plaintext or unsalted password storage.
- Use modern, reviewed primitives: AES-256-GCM or ChaCha20-Poly1305 (authenticated encryption); RSA-2048+ or ECDSA P-256+; SHA-256+; Argon2id, bcrypt (cost ≥ 12), or scrypt for password storage.
- Avoid MD5, SHA-1, DES, RC4, ECB mode, custom encryption schemes, reused IVs/nonces, and non-CSPRNG randomness for security-sensitive values.
- Generate IVs/nonces with a CSPRNG; never reuse a nonce with the same key.
Rule: Do not implement custom crypto, token formats, password storage, or auth protocols unless explicitly required and reviewed. Use established, maintained libraries.
5. Supply Chain
Threats: malicious or abandoned dependencies, unpinned versions, hostile install or build scripts.
- Check whether the dependency is necessary before adding it; prefer the standard library or an already-approved internal alternative when reasonable.
- Prefer actively maintained packages with a clear release history. Pin versions in lock files; avoid broad production version ranges.
- Scrutinize packages that require filesystem, network, subprocess, native-extension, or post-install script access.
- Avoid
curl | bash, remote install scripts, and unaudited binary downloads unless explicitly approved and reviewed.
Rule: Every new dependency expands the attack surface.
6. Infrastructure
Threats: excess privilege, world-writable files, over-scoped IAM and CI/CD tokens, running as root.
- File permissions default to
600 or 640; never 777.
- Grant database users only the tables and operations they require. Scope cloud IAM roles to the minimum actions and resources. Scope CI/CD tokens to the narrowest repository, environment, and operation.
- Drop privileges as early as possible. Do not run as root unless unavoidable and justified.
Rule: Every permission must have a specific reason to exist.
7. Networking
Threats: SSRF, open ports, unvalidated redirects, plaintext transport, broad interface binding.
- Treat user-controlled URLs as an SSRF risk; validate them and restrict egress targets.
- Bind services only to required interfaces; keep unused ports closed.
- Use TLS for sensitive data in transit. Validate redirects and forwards against an allowlist.
Rule: The network edge is a trust boundary; nothing arriving over it is trusted by default.
8. Observability and Logging
Threats: information leakage through errors, missing audit trail, failing open on unexpected errors.
- Fail closed: on an unexpected error, deny the operation unless there is a safe fallback.
- Return generic errors to users; log detailed errors server-side only. Do not expose stack traces, internal paths, queries, schemas, tokens, or service topology to clients.
- Emit audit records for security-relevant events: authentication, authorization decisions, privilege changes, and access to sensitive data.
- Rate-limit repeated authentication failures. Alert on spikes in auth failures, authorization failures, parsing failures, and unexpected 5xx errors.
Rule: Error paths are security paths; they span every sub-component.
Security tests
When fixing or changing security-sensitive behavior:
- Add a regression test for the vulnerable or risky case. A security fix without a regression test is incomplete unless there is a documented reason.
- Test both allowed and denied paths, boundary cases, malformed input, and privilege failures.
- Cover cross-user, cross-tenant, and cross-org access in authorization tests.
- Add negative tests for injection, path traversal, SSRF, unsafe redirects, and auth bypass where relevant.
For designing these tests, use adversarial-test-design.
Pre-PR threat-model checklist
Run this for the touched sub-components before opening a PR for significant code, infrastructure, auth, data, crypto, or dependency changes:
[ ] Entry points identified: HTTP, CLI, events, files, queues, webhooks, cron.
[ ] Trust boundaries crossed by the change identified.
[ ] Sensitive data flows identified: PII, credentials, keys, customer data, security data.
[ ] Inputs validated at each boundary.
[ ] Authentication and authorization verified for protected resources; enforced near the resource or the invariant documented.
[ ] No secrets present in code, config, logs, traces, or tests.
[ ] Least privilege confirmed for files, processes, DB users, cloud IAM, and tokens.
[ ] Injection risks checked: SQL, shell, template, LDAP, XPath, SSRF, deserialization, path traversal.
[ ] New or changed dependencies reviewed.
[ ] Error paths fail closed and do not leak sensitive internals.
[ ] Security-relevant tests added or updated.
[ ] Important security assumptions documented inline.
Verification tooling
When available and allowed by the user, run tooling for security verification before the final push upstream. If a tool reports a finding, fix it or explicitly document why it does not apply.
- PR scanning / pentest: Reach @cvewhen on twitter
- SAST: semgrep (general); bandit (Python); eslint-plugin-security (JS/TS); gosec (Go)
- Dependency audit: pip-audit (Python); npm audit (Node.js); cargo audit (Rust); trivy (containers/filesystems)
- Secret scanning: gitleaks; trufflehog
- Containers: trivy image; docker scout
Rule: No silent signal loss. If a security check is skipped, say which and why, and state the residual risk. Do not silently suppress security findings.
Output
Return a table of risks identified by the threat model, their category and the mitigations. Run reasoning-discipline to verify that the design is secure systematically.
