Find missing coverage, unfinished triage and structural issues. Work through the findings, save a corrected model, and give your reviewer a clearer starting point.
This tool improves the threat model. It does not patch application code, prove a system secure or guarantee Microsoft approval.
Turn model gaps into reviewable work.
Use it before a review, while drafting a model, or when checking what changed.
Know what needs attention
Open a .tm7 or Threat Dragon .json. Get a score, readiness verdict and findings with check IDs and references, rather than an unexplained AI judgement.
Correct the model deliberately
Preview missing threats, triage changes and draft justifications. Choose what to apply and save a new .tm7. A proposed mitigation still needs an owner and verification.
Make the hand-off actionable
Export reports or tracker-ready work items for GitHub, Azure DevOps and Jira. Compare revisions and keep local review history. Optional AI can help explain findings and propose mitigation tests.
Start without a blank diagram
Draft from a description, supported architecture files or an Azure resource group. Review the inferred components and paths before generating threats. Use a suitable TMT baseline when you need to reopen newly authored files in Microsoft TMT.
Azure is an additional starting point, not a complete traffic map.Read-only metadata queries and eligible role assignments suggest possible access paths. Shared-key access and out-of-scope dependencies can be missing; review the evidence and the draft.
Try it before opening your own model.
The synthetic customer portal has intentionally unfinished triage. Open it in the app, inspect the C1 findings, then preview a fix plan. No Copilot sign-in is needed.
Expected: NOT READY, 65/100, with 27 gating findings. Exit code 2 means the review found blockers; 1 means the command could not run.
Moving a threat to Needs Investigation is not a security fix. Keep investigating and record implementation evidence.
What it looks like
A real review of the synthetic customer portal. The sample result is reproducible under the default rubric. Screenshots use a development build; small UI details may differ from the released installer.
Overview. The verdict and the score, and the four weighted parts the score is made of. Readiness is a separate gate from maturity — a thorough model can still be blocked by one must-fix issue, and the panel says which.
Findings. Every issue carries the id of the check that raised it, so a disagreement is about a rule rather than about a judgement.Fix. A remediation plan you review before anything is written. Each row says whether it came from a built-in rule or from Copilot, and the result is a valid .tm7 that reopens in the Threat Modeling Tool.Diagram. The data-flow diagram, with real Microsoft Threat Modeling Tool stencil artwork when that tool is installed. Drag to rearrange; edits save back to the file.
How the verdict is produced
The verdict is computed before any optional AI explanation.
1
Parse
The .tm7 is read into elements, flows, trust boundaries and threats. OWASP Threat Dragon .json works too, so a mixed toolchain lands in one place.
2
Evaluate
72 rules run over that structure. Each is ordinary code with a fixed identifier — C1, A6, L3 — so a finding can be traced to the exact rule that raised it and argued with on its merits.
3
Score and gate
Maturity is a weighted 0–100 score. Readiness is a separate gate: a high-scoring model can still have a must-fix finding. These are signals about the model, not proof that controls are deployed.
4
Explain, optionally
Only now can Copilot help — describing a finding, drafting a mitigation, critiquing the model. It cannot change the score or the verdict, because those were already computed.
The same model, rubric version and policy produce the same result.
READY WITH NOTES means no gating findings remain under that configuration.
Your reviewer still decides what evidence and risk treatment are sufficient.
What the review checks
72 deterministic checks apply public security guidance to the model you provide. Each finding can be traced to its rule.
Structure and scope
Missing endpoints, orphaned elements, ambiguous boundaries and out-of-scope components without a reason.
STRIDE coverage
Missing categories and unanalysed interactions, including flows crossing trust boundaries.
Triage and justification
Threats still Not Started, unsupported Mitigated or Not Applicable claims, and boilerplate explanations.
AI and agentic risks
Applicable OWASP LLM Top 10 (2025) and MITRE ATLAS risks, including prompt injection, retrieval exposure and excessive agency.
Security hygiene
Model signals for identity, data protection, logging, deprecated cryptography and related security concerns. These checks do not scan deployed code.
Coverage evidence
STRIDE and framework scorecards, check IDs and source references. Use them to discuss gaps, not as a compliance certificate.
Use it where you work.
The same deterministic review engine, with a surface suited to your task.
Desktop app
Inspect diagrams, review findings and preview model edits. Create and Assistant support drafting; Ask and floating Ask share local and optional AI modes.
Run reviews in CI, export SARIF, compare revisions or apply an organization policy with disclosure. The headless bundle is separate from the desktop installer.
Ask an agent to invoke the reviewer and report its actual results. Install the skill and the CLI bundle together; the instructions do not replace the engine.
Ask an agent to review, explain, fix or compare a model in plain English. The verdict still comes from the deterministic engine — the agent only runs it. Needs the command-line bundle alongside it.
On v1.0.4 or later the app updates itself — click Download & install on the in-app banner, then Install & restart. Every version, its notes and all files are on the releases page.
Optional developer context uses eight restricted operations across Microsoft Learn, Azure metadata and GitHub. Sources default to off, require consent, and do not affect the deterministic verdict. Read the MCP setup guide and Microsoft TMT compatibility limits.
SmartScreen note: executables and installers use a self-signed certificate, so Windows may warn or block them. Verify the download and follow your organization's policy. A future trusted-publisher certificate does not guarantee that every warning disappears.
The score is yours to trust
The parts that decide readiness are fully deterministic and run locally. Copilot only assists.
Deterministic verdict
The NOT READY / READY-WITH-NOTES verdict, the 0–100 score and the findings come from a fixed rubric engine — never from AI. Copilot only explains findings and drafts fixes you choose to apply, so the same model always scores the same way.
Private by design
Deterministic review runs locally. Optional AI actions send context to your chosen provider; image extraction sends the selected image. Read the privacy guide before using confidential inputs.
No telemetry
The app does not collect product analytics. Update checks can be disabled. AI, Azure discovery and enabled MCP sources are separate network features; external runtimes have their own settings.
Built for enterprise rollout
Everything a security review and a deployment team need, documented up front.
✓ Security & privacy review
Every network destination, on-disk storage path and redaction rule is enumerated — plus how to verify the claims yourself with a proxy or offline run.
✓ Silent, managed installation
MSI with per-machine or per-user scope, documented UpgradeCode and silent switches for Intune, SCCM and Group Policy.
✓ Offline review
Disable update checks and leave remote AI, Azure discovery and MCP sources unused. Deterministic review and built-in model fixes need no network.
✓ Keep inference in your tenant
Use GitHub Copilot, or point the provider at Azure OpenAI or a self-hosted OpenAI-compatible endpoint. Credentials are DPAPI-encrypted locally.
✓ CI/CD gating
CLI review runs the deterministic engine headlessly and exits 2 for NOT READY. --sarif writes a local file; upload to code scanning is a separate workflow.
✓ Signing information
Executables and installers use the project's current self-signed certificate. ZIP containers are not Authenticode-signed. Follow the install guide and your organization's software policy.
