security-sweeplisted
Install: claude install-skill samiraklf/claudrunner
# Security sweep
Find the holes. Report only what you can prove from the code.
## Scope
Given a diff, review the diff and everything it reaches. Given a codebase, start at the
boundaries: every route, handler, message consumer, scheduled job, webhook and file upload.
## What to hunt
**Injection.** Any query, command, path, template or markup assembled from input that a
user influences. Parameter binding is the fix; escaping is not. Look for string
concatenation into a query builder as often as raw statements.
**Authorization.** Every endpoint answers three questions: is the caller authenticated, are
they permitted to do this, and are they permitted to do it *to this object*. The third is
the one that gets missed. An identifier accepted from a URL or body, then used without an
ownership check, is the most common real breach in a business application.
**Tenant isolation.** In any system with more than one customer, every read and write on
customer-owned data must be scoped to the caller's tenant, derived from the session — never
from a parameter the caller supplies.
**Secrets.** Credentials, tokens and keys in source, in fixtures, in test files, in logs, in
error messages, in client-side bundles. Check what the logger records on failure.
**Mass assignment.** Any path where a whole request body becomes a record's fields.
**Deserialization and templating.** Untrusted input reaching a deserializer, a template
engine, an expression evaluator, or a dynamic import.
**Reque