Security Regression Guard

Why it exists

Hiding a button does not stop a direct request. This skill takes the agent from product policy to the routes, queries, writes, and tests that actually protect an operation.

Its scope stays proportional to the change: it reviews sensitive boundaries without turning every visual task into a full audit or replacing professional security review.

What it reviews

• Roles, permissions, and alternate paths to the same action.
• Isolation by person, organization, and resource ownership.
• Sessions, files, and operations with partial effects.
• Limits before expensive work begins.
• Direct unauthorized requests and their lack of side effects.
• The allowed case after the correction is applied.

A real shift-rotation case

During the implementation of a multi-tenant shift-rotation module, an initial functional version passed its positive cases but left four boundaries uncovered: a missing read permission, cross-organization enumeration, unbounded input, and an out-of-range starting index.

Reviewing the change with the skill turned those omissions into specific corrections and eleven tests covering both side-effect-free rejection and the authorized flow.

The skill provides reusable criteria. It does not add a security barrier by itself or guarantee that a model will apply every instruction correctly.