Turn a code diff into a focused authorization test plan

Falsifiable hypothesis

A diff-focused assistant can expose overlooked authorization test cases when every suggested case must cite both the changed code and the relevant enforcement path.

Source-led motivation

Authorized lab scope

A trusted copy of an application you own or are authorized to assess, with a disposable staging deployment and synthetic users in distinct roles.

What to try

• Give the assistant the trusted diff, relevant route handlers, authorization middleware, and a short description of intended roles. Treat comments and repository text as data, not permission to run tools.
• Ask for a table: changed behavior, caller-controlled input, required role or ownership rule, enforcement location, and source file/line evidence. Require “unknown” when context is missing.
• Ask it to propose positive and negative tests for each supported boundary: a permitted user succeeds, a different synthetic user is denied, and unauthenticated access follows the intended policy. Do not infer a vulnerability from a missing check in one file.
• Review the candidates, account for shared middleware, then run approved cases against staging. Keep minimal request/response evidence and classify each lead as confirmed, rejected, or unresolved.

Prerequisites

• A trusted code revision and intended authorization policy.
• Synthetic owner, non-owner, and lower-privilege accounts; known-good control cases.
• A model route approved for this code and no production credentials in the analysis context.

Comparison & control

Have an analyst produce an initial test list before seeing the assistant output. Compare additional valid cases, false leads caused by omitted middleware, and time to verification; keep both lists.

Observations to collect

• Source locations supporting each suggested boundary.
• Expected versus observed allow/deny decisions, with synthetic record ownership documented.
• Human-added cases, hallucinated handlers, and mistaken vulnerability claims.

Disconfirming evidence

• The assistant adds no valid cases or repeatedly mistakes centralized authorization for absent authorization.

Stop conditions

• Stop at the approved target and data boundaries; do not send client artifacts to a model provider without approval.
• Stop before destructive operations or unexpected effects; reproduce only with disposable accounts and synthetic records.

Source bibliography

1. Claude Code Security Reviewer ↗

   Anthropic / anthropics

   Publication date: Date not supplied · Editorial check 2026-10-26 UTC

   Project README / official documentation; source-led editorial check, no execution

Continue the thread

• AI code review has a trust boundary of its own →