Review security

Command: /security-flow · ← All scenarios · User guide

Run an authorized, evidence-preserving security review that ends with sanitized findings and concise remediation-task inputs. It reviews and reports — it does not fix anything or start coding.

Use this when you need an authorized security review of an application, repository, infrastructure, interface, host, or AI system — including PR and pipeline reviews or a broad threat-model-driven pass.

Not for: fixing findings (it hands off task inputs for a later coding session) or any unauthorized/production active testing.

Running it

/security-flow Review this service and its infrastructure. Recommend a safe full-review scope and wait for approval before inspection.
/security-flow Review the payment-service PR on this branch, read-only, pre-production only.
/security-flow Threat-model the checkout API and run all applicable authorized read-only checks.

How it works

Safety is built into the order of operations. Before any source is read, a secret gate scans filenames only. The agent proposes a scope you must approve, does the review within those bounds, has an independent reviewer challenge it, and packages sanitized results.

flowchart TB
    Start(["/security-flow + authorized scope"]) --> Ready["Readiness + secret gate <br>(filenames only)"]
    Ready --> SG{"Secret-gate <br>result?"}
    SG -->|stop / high-risk| Halt(["Review halts — <br>cannot be overridden"])
    SG -->|needs approval| Auth["Recommend run scope"]
    SG -->|pass| Auth
    Auth --> G1{"You approve <br>the run scope"}
    G1 -->|adjust| Auth
    G1 -->|approve| DG{"Development, change, <br>PR or pipeline review?"}
    DG -->|yes| Gates["Run deterministic gates"]
    DG -->|no| Model["Threat model + coverage plan"]
    Gates --> GR{"High+ findings?"}
    GR -->|yes| Package["Report & package"]
    GR -->|no| Model
    Model --> Inspect["Inspect & test <br>(pre-production only)"]
    Inspect --> Triage["Normalize & triage findings"]
    Triage --> Indep["Independent reviewer"]
    Indep --> IR{"Review <br>accepted?"}
    IR -->|corrections| Triage
    IR -->|accepted| Package
    Package --> G2{"You approve <br>the task index"}
    G2 -->|request changes| Package
    G2 -->|approve| Done(["Sanitized report + remediation tasks"])

    classDef step fill:#dae8ff,stroke:#3674b5,color:#102a43;
    classDef gate fill:#fff2cc,stroke:#b58b00,color:#493800;
    classDef done fill:#e5f7e8,stroke:#35834a,color:#153b20;
    classDef halt fill:#ffe0e0,stroke:#b54040,color:#5a1a1a;
    class Ready,Auth,Gates,Model,Inspect,Triage,Indep,Package step;
    class SG,G1,DG,GR,IR,G2 gate;
    class Start,Done done;
    class Halt halt;

The guardrails, plainly

What you’ll be asked to do

Provide the authorized scope and environment; approve the run contract (activities, tools, data flows, stop conditions, how much exploit detail the report carries); approve any flagged DEV/QA candidate files; and approve the remediation task index at the end. The agent won’t commit or delete on your behalf — you review and commit the artifacts.

What it creates

With your storage approval, sanitized artifacts under docs/security/<run-id>/: report.md, findings.json, run.json, a tasks/INDEX.md, and per-group tasks/<task-id>.md files (the inputs for later remediation). Raw scanner output stays local and is never committed.

Write or change code to act on the remediation tasks afterward.

Sources