threat-model-gate

Threat Model Gate

Overview

Security cannot be added later. Features that touch user data, system boundaries, or trust relationships require threat modeling during planning — not after code is written.

Core Principle

Every feature that handles user data, authentication, or system boundaries requires explicit threat modeling before implementation.

"We'll secure it in the next sprint" is a commitment that never gets honored. The cost of retrofitting security is 10x the cost of designing it in. Threat modeling during planning is the minimum viable security practice.

When to Activate

Apply this gate when planning involves any of the following:

Authentication and Identity

• Login flows, session management, token issuance or validation
• Multi-factor authentication, SSO, OAuth, OIDC, SAML
• Password storage, credential management, account recovery

Authorization and Access Control

• Role-based or attribute-based access control (RBAC, ABAC)
• Resource ownership and sharing models
• Admin capabilities, privilege escalation paths
• API key or service account permissions

Data Handling

• PII collection, storage, or transmission
• Payment card data, health records, or regulated data categories
• Encryption at rest or in transit
• Data retention, deletion, or export features

API Design

• Public-facing endpoints, webhooks, or callbacks
• Inter-service communication with trust implications
• Rate limiting and abuse prevention
• Input validation and output encoding boundaries

Cryptography

• Key generation, storage, or rotation
• Algorithm selection (hashing, signing, encryption)
• Certificate management, mTLS
• Secure random number generation

Network and Infrastructure

• Firewall rules, network segmentation, VPN configurations
• Load balancer, reverse proxy, or CDN configurations
• Container networking, service mesh, Kubernetes network policies
• DNS, TLS termination, certificate pinning

The Threat Modeling Checklist

Work through these four areas before committing to an implementation approach:

1. Assets

Identify what is worth protecting:

ASSETS TO IDENTIFY:
- What data does this feature create, read, update, or delete?
- What is the confidentiality classification? (public / internal / confidential / restricted)
- What is the integrity requirement? (can corruption be tolerated? for how long?)
- What is the availability requirement? (what is the acceptable downtime?)
- What are the downstream systems or users that depend on this data?

2. Trust Boundaries

Map where trust changes:

BOUNDARIES TO IDENTIFY:
- Where does data cross from an untrusted zone to a trusted zone?
- Which actors (users, services, admins) receive which level of trust?
- Where are authentication and authorization enforced?
- What can an unauthenticated caller reach?
- What can an authenticated-but-unauthorized caller reach?
- Where does input become data (the injection boundary)?

3. Threats (STRIDE Analysis)

Apply STRIDE to each asset and boundary (see STRIDE Quick Reference below):

FOR EACH BOUNDARY OR ASSET:
- S: How could an attacker impersonate a legitimate actor?
- T: How could an attacker modify data in transit or at rest?
- R: How could an actor deny having performed an action?
- I: How could an attacker read data they should not see?
- D: How could an attacker make the feature unavailable?
- E: How could an attacker gain more privilege than intended?

4. Mitigations

For each threat identified, specify the control:

MITIGATIONS TO SPECIFY:
- Authentication controls (how is identity verified?)
- Authorization controls (how is permission verified?)
- Input validation (what are the trust boundaries for input?)
- Encryption (what is encrypted, with which algorithm, where?)
- Audit logging (what events are logged, where, for how long?)
- Rate limiting (what abuse scenarios does this prevent?)
- Failure mode (what happens when the control fails?)

STRIDE Quick Reference

STRIDE is a starting point, not a complete threat model. Use it to ensure you have covered all six threat categories, then go deeper on the categories most relevant to the feature.

DREAD Risk Rating Reference

When threat modeling identifies specific threats, score each using DREAD (5 dimensions, each 0-10). The average determines severity classification.

Severity bands:

• CRITICAL: DREAD average >= 8.0
• HIGH: DREAD average >= 6.0 and < 8.0
• MEDIUM: DREAD average >= 4.0 and < 6.0
• LOW: DREAD average < 4.0

Calibration rule: When a score falls within 0.5 of a boundary (e.g., 5.5-6.4), check Damage Potential. If DP >= 7, round UP. If DP <= 3, round DOWN.

Security Requirements Template for Plans

Every plan for a security-sensitive feature must include a ## Security Requirements section:

## Security Requirements

### Assets
- **[Asset Name]:** [Confidentiality: public/internal/confidential/restricted] | [Integrity: high/medium/low] | [Availability: high/medium/low]

### Trust Boundaries
- **Boundary:** [Description — e.g., "public internet to application server"]
  - **Authentication:** [How identity is established]
  - **Authorization:** [How permission is verified]

### STRIDE Analysis
| Threat | Vector | Mitigation | Residual Risk |
|--------|--------|-----------|---------------|
| Spoofing | [How an attacker could spoof] | [Control] | [low/medium/high] |
| Tampering | [How data could be tampered with] | [Control] | [low/medium/high] |
| Repudiation | [What actions could be denied] | [Control] | [low/medium/high] |
| Information Disclosure | [What data could be exposed] | [Control] | [low/medium/high] |
| Denial of Service | [What could be exhausted or crashed] | [Control] | [low/medium/high] |
| Elevation of Privilege | [How privilege could be escalated] | [Control] | [low/medium/high] |

### Security Controls
- **Input Validation:** [What is validated, where, and how]
- **Encryption:** [At rest: algorithm. In transit: TLS version, cipher suites]
- **Audit Logging:** [What events are logged, retention period, tamper protection]
- **Rate Limiting:** [Limits, scope, response on breach]
- **Secrets Management:** [Where credentials are stored, how rotated]

### Failure Modes
- **If authentication fails:** [Behavior — e.g., "return 401, log attempt, no information leakage"]
- **If authorization fails:** [Behavior — e.g., "return 403, log with user ID and resource"]
- **If encryption fails:** [Behavior — e.g., "abort operation, do not fall back to plaintext"]

Anti-Patterns

When you notice these in a plan, stop and apply threat modeling before proceeding:

"Security will be added later"

WRONG: "We'll add authentication in v2."
RIGHT: Define the authentication model now. Implementation can be phased; the design cannot.

Implicit trust of internal services

WRONG: "It's only called by our internal API, so we don't need auth."
RIGHT: Internal services are compromised too. Define what trust means and how it is enforced.

Encryption as an afterthought

WRONG: "We'll encrypt the database later when we have time."
RIGHT: Define encryption at rest requirements now. Schema changes after launch are expensive.

Authorization by obscurity

WRONG: "Users won't know the endpoint exists."
RIGHT: Assume all endpoints are discoverable. Enforce authorization explicitly.

Logging as a security control

WRONG: "We'll know if something bad happens because we log everything."
RIGHT: Logging is detection and response, not prevention. Identify the preventive controls.

"We trust our users"

WRONG: "Our users are internal employees, they wouldn't abuse this."
RIGHT: Insider threat is real. Least privilege applies to employees too. Define what each role can do.

Deferring the threat model to the security team

WRONG: "Security will review this before it ships."
RIGHT: The security team reviews your threat model. You write the threat model. Start now.

Sharing secrets in plans or commits

WRONG: Including actual API keys, connection strings, or credentials in plan documents or code.
RIGHT: Reference secrets by name only. Actual values belong in secrets managers, not in plans.

9. Incomplete trust boundary analysis

   • WRONG: "The API talks to the database" (no trust boundary identified)
   • RIGHT: "Data flow DF-003 crosses from TZ-DMZ to TZ-Internal at the API-to-DB boundary, carrying user credentials over PostgreSQL wire protocol"

10. Skipping DREAD calibration

    • WRONG: DREAD average 5.8, classified as MEDIUM
    • RIGHT: DREAD average 5.8 with Damage Potential 8 -> calibrate UP to HIGH (DP >= 7 near boundary)

How to Apply During Planning

When drafting or reviewing a plan for a security-sensitive feature:

1. Identify the security scope. Does this plan touch any of the activation categories listed above? If yes, the threat model gate is active.

2. Audit the plan for Security Requirements. Is there a ## Security Requirements section with assets, trust boundaries, STRIDE analysis, and security controls documented? If not, the plan is incomplete.

3. Apply STRIDE. Walk through each threat category for each identified asset and boundary. If any category has no entry, ask: "Have we genuinely considered this, or have we overlooked it?"

4. Verify mitigations are specific. "We'll use standard security practices" is not a mitigation. "Passwords hashed with bcrypt (cost factor 12), stored in the users table password_hash column, never logged" is a mitigation.

5. Check failure modes. What does the feature do when a security control fails? Graceful failure is part of the design.

6. Flag anti-patterns. If the plan contains any of the anti-patterns above, surface them explicitly before the plan is approved.

Relationship to Other Skills

• Before /architect approval: The threat model gate runs during planning. Plans without a ## Security Requirements section for security-sensitive features should not receive architect approval.
• Before /ship implementation: If a plan passed /architect but the Security Requirements section is missing or shallow, raise it before implementation begins. It is cheaper to fix the design than the code.
• After /secure-review: If /secure-review finds issues that the threat model should have caught, update the threat model section of the plan and the .claude/learnings.md file. Threat model gaps are learning opportunities.
• In conjunction with receiving-code-review: When reviewing security-related code, apply both skills. receiving-code-review governs how you evaluate the reviewer's feedback; threat-model-gate governs whether the implementation addresses the correct threats.

Full Threat Modeling Methodology

This skill provides a planning gate -- it ensures threat modeling happens before implementation. For the full threat modeling methodology with three-phase STRIDE+DREAD analysis and OTM JSON output, load the standalone threat-model knowledge-base skill:

Using skills/threat-model/SKILL.md: perform a threat model for this system.

The threat-model skill provides:

• Three-phase workflow: System Decomposition, Per-Subsystem Threat Analysis, Cross-Cutting Synthesis
• DREAD scoring with calibration and consistency checks
• MITRE ATT&CK technique mapping
• OTM v0.2.0 JSON output (machine-readable)
• Structured markdown report (human-readable)
• Reference materials in reference/otm-schema.md and reference/report-template.md

The Bottom Line

Security is a design constraint, not a feature. It cannot be sprinted in after the architecture is set.

Threat modeling during planning takes 30 minutes. Retrofitting security after launch takes months and may require breaking API changes, data migrations, or architectural rewrites.

Model the threats. Document the controls. Then build.