Threat Model

threat-model

Finds the attacks a design permits before the code exists, when fixes are a diagram edit instead of a migration. Design-phase counterpart to devils-advocate (which attacks written code); consumes design-doc or reverse-architecture output directly. Threats live where data crosses a trust boundary — the model is anchored there, not in a generic checklist.

How to respond

1. Establish the system before any threats. From the input, extract and state:

• Assets — what an attacker wants (credentials, PII, payment data, compute, availability), ranked by damage-if-lost
• Entry points — every way data or commands get in (APIs, queues, file uploads, admin panels, CI/CD, third-party webhooks)
• Trust boundaries — every line where the trust level changes (internet→edge, service→service, app→DB, human→system, tenant→tenant)

Ask at most 2 questions, and spend them on trust boundaries — misplaced boundaries invalidate the whole model. Tag everything else [assumption].

1. Enumerate boundary crossings, then run STRIDE per crossing using the question banks in [reference.md](reference.md) — Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, Elevation of privilege. Not every category applies at every crossing; say "n/a" with one word of why rather than padding.

1. Write each threat as a concrete attack scenario: actor → path → result. ✅ "A tenant crafts a report name containing ../, the export service writes to another tenant's S3 prefix, cross-tenant data exposure" — ❌ "Path traversal is possible". A threat that can't be narrated as a scenario isn't understood yet.

1. Severity-rank with the shared markers — 🟥 critical (unauthenticated or cross-tenant compromise of a top asset), 🟧 high (authenticated bypass, single-tenant data exposure), 🟨 medium (requires chaining or insider position), ⚪ low/accepted. Likelihood × impact, judged against this system's assets — not CVSS theater. If everything is 🟥, the ranking failed.

1. Assess existing mitigations honestly before recommending new ones. Each threat gets: what already mitigates it (and whether that's verified or assumed), then the recommended mitigation with a verification step — ✅ "Presigned URLs scoped to tenant prefix; verify with a cross-tenant fetch test in CI" — ❌ "Use proper access controls".

1. Force the insider and authz passes. Two systematic blind spots, swept explicitly: what can a legitimate authenticated user do to another tenant or to escalate (authn is not authz), and what can a malicious or compromised insider/CI-token do. One finding or an explicit all-clear for each.

1. Emit the model in one message: system summary (assets / entry points / boundaries), findings grouped by severity, the n/a ledger, assumptions, out of scope, and questions for the team — the unknowns that block a finding from being confirmed or dismissed. End with the revisit trigger: the model is stale when a new entry point, boundary, or asset class appears.

Useful references in this skill

• [reference.md](reference.md) — STRIDE question banks per boundary type + common-mitigation table
• [examples/document-upload.md](examples/document-upload.md) — worked example: upload feature → 6 ranked findings

Quality bar

• Every threat is anchored to a named trust-boundary crossing or asset. An unanchored finding is checklist residue — cut it.
• Every threat narrates a scenario with actor, path, and result. "X is possible" is not a finding.
• Severity spread is honest. A model where everything is critical (or nothing is) hasn't ranked anything; the 🟥 set is what you'd halt the build for.
• Every recommended mitigation has a verification step a reviewer can run or check. Advice without verification is a wish.
• Existing mitigations are credited — and marked verified vs assumed. A model that ignores what the design already does right teaches the team to ignore the model.
• The insider and authz sweeps produced output — findings or an explicit reasoned all-clear, never silence.

When to use this skill

• ✅ A design doc or RFC needs security scrutiny before implementation starts
• ✅ "How could this be attacked?" for a new feature, integration, or entry point
• ✅ Input for a formal security review / sign-off gate
• ✅ A reverse-architecture extraction of a legacy system that never had a threat model

When NOT to use this skill

• ❌ Reviewing written code for vulnerabilities — that's devils-advocate (its security lens) or sonar-review
• ❌ Compliance attestation (SOC2, PCI evidence) — this models attacks, it doesn't certify controls
• ❌ Incident analysis after an exploit — that's postmortem-generator
• ❌ Pen-testing a live system — this is paper analysis; it tells the pen test where to aim

Anti-patterns to avoid

• ❌ OWASP-checklist dumping — twenty generic items ("ensure input validation") untied to any boundary in this system. Volume isn't rigor.
• ❌ Stopping at authentication. "Users must log in" answers spoofing only; the authz pass (tenant→tenant, user→admin) is where real findings live.
• ❌ "Use encryption" as a mitigation. Of what, where, against which threat? Encryption at rest does nothing for a tampered request in flight.
• ❌ Modeling only the happy architecture. Admin panels, CI/CD pipelines, support tooling, and webhook receivers are entry points — usually the softest ones.
• ❌ Inflating severity to be taken seriously. One cried-wolf 🟥 and the next model gets skimmed.
• ❌ One-and-done modeling. A threat model without a revisit trigger expires silently the first time the design changes.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sananthanarayan
• Source: sananthanarayan/skilldrop
• License: MIT

Install and usage instructions live in the source repository linked above.