Security architecture review · grounded in source
Follow the data.
Find the boundary.
Turn a real repository into a concise threat model: the assets worth protecting, the paths an attacker could take, and the controls that should close them. Every architectural claim is tied to a path in your code.
Where can an untrusted choice cross into a trusted action?
That is the question this Skill keeps asking. It starts at entry points, traces the code, and writes down only the components, flows, and controls it can point to in the repository. Anything it cannot see becomes a stated assumption.
What it maps
This illustration groups typical inputs, application logic and protected assets; the actual system model follows your repository. The dashed edges are the trust boundaries. Threats are written as abuse paths that cross them.
Zone 1
Untrusted input
- Endpoints and RPC handlers
- Upload surfaces and parsers
- Job triggers and webhooks
- CLI arguments, files, env
Zone 2
Application logic
- Authentication and sessions
- Authorization, RBAC, ACLs
- Validation and deserialization
- Outbound requests, subprocesses
Zone 3
Protected assets
- Credentials, keys, tokens
- PII and tenant data
- Config and integrity state
- Build artifacts, audit logs
You supply. You receive.
Supply
- Repository accessThe repo root, or the specific paths that are in scope.
- Service contextIntended usage, deployment model, internet exposure, and authentication and authorization expectations, as far as you know them.
- Sensitivity and limitsData sensitivity, tenancy, and what is out of scope.
- A short confirmationAnswers to one to three targeted questions before the final report is written.
Receive
<repo-or-dir>-threat-model.mdOne Markdown report in a fixed section order.- System model and diagramComponents, data flows, trust boundaries, and a compact Mermaid flowchart.
- Assets, attackers, entry pointsTables of assets with C/I/A objectives, attacker capabilities and non-capabilities, and evidence-anchored surfaces.
- Prioritized threatsAbuse paths, a TM-001… threat table with likelihood, impact, and priority, specific mitigations, and focus paths for manual review.
When to engage it
Engage for
- AppSec engineers who need a repository-specific threat model, not a generic checklist.
- Developers preparing a codebase or a single path for security review.
- Teams examining a sensitive boundary such as tenancy, authentication, uploads, or outbound fetches.
- Explicit requests to enumerate threats, abuse paths, and mitigations.
Look elsewhere for
- General architecture summaries or non-security design work. The Skill is written not to trigger for these.
- Routine code review of a change.
- Automated vulnerability scanning. No scanning engine is bundled.
- A guarantee of safety. Finding nothing does not prove there is nothing to find.
Operating requirements
- Assistant
- A coding assistant that loads Agent Skills (a
SKILL.mdplus reference files) and can read your repository. The package includes Codex interface metadata inagents/openai.yaml. Other assistants were not tested separately. - Tools
- File reading and code search. The reference template suggests ripgrep with
-Iwhen it is available. Nothing else is installed or bundled. - API keys
- None extraThe Skill calls no separate service, so it needs no service API key.
- Model access
- Still requiredThe assistant itself needs model access, and any usage costs of that assistant still apply.
- Permissions
- You manage access to the repository and any infrastructure. Review only code and data you are authorized to inspect, keep secrets out of reports (the template says to redact them), and run any runtime reproduction only in an approved environment.
What has been verified
Inside the evidence
- All five original upstream files match their recorded SHA-256 values.
- One synthetic two-tenant document API was reviewed from source, and its deliberate authorization defect was reproduced with three localhost requests.
- The agent-authored sample report contains every required heading from the output contract.
Outside the evidence
- Interactive context confirmation with a real service owner.
- Production infrastructure, real identities, and non-synthetic data.
- Detection accuracy, comparative scores, and validation of a fix.
This is a reasoning workflow, not a scanner. Results depend on the assistant, the source it can read, the context you give it, and how carefully the report is reviewed. The Evidence sheet sets out exactly what the exercise establishes.
Provenance
- Origin
- Original OpenAI Skill
security-threat-modelfrom the openai/skills collection, pathskills/.curated/security-threat-model. - Pinned revision
- 49f948faa9258a0c61caceaf225e179651397431
- Preserved files
SKILL.md,agents/openai.yaml,references/prompt-template.md,references/security-controls-and-assets.md, andLICENSE.txt, byte for byte. Added for packaging only:PROVENANCE.jsonandREADME.md.- License
- Apache-2.0, included unchanged.
- Status
- The upstream openai/skills collection is deprecated and points new examples to openai/plugins. This is a fixed historical distribution of the official Skill. It is not a current plugin release, an official vendor storefront, or an endorsement.