Public Client Leak Audit

Audit a public client codebase that talks to a private backend. Goal: the public surface must be self-contained — it may reveal the calls it makes and the data shapes it exchanges (unavoidable for any shipped client), but nothing beyond that. Every extra detail about how the server works is free reconnaissance for an attacker and a lever for abuse.

This skill produces three things: a findings list (leaks + client-side security holes, each with file:line and severity), a set of applied fixes, and a report with residual recommendations.

Reference files (load on demand — read the one you need, don't inline all of them):

• references/leak-taxonomy.md — the categories of disclosure to hunt, why each matters, and starter search patterns. Read it before the first sweep of the repository.
• references/rewrite-rules.md — the comment/string rewrite rule with before/after examples; how to decide keep-vs-cut. Read it when a finding is about to be rewritten, not while hunting.
• references/client-hardening.md — runtime-independent client-side security checklist (capabilities, cross-context entry points, tokens, network, build config, supply chain). Read it at the hardening phase, whatever the client type.
• references/browser-client.md — the browser half of that checklist (extension permissions, storage tiers, DOM/CSS sinks, bundler config, npm lifecycle scripts). Load it with client-hardening.md for an extension, SPA, or web SDK; skip it for a native, desktop, or CLI client.
• references/report-template.md — the output report structure. Read it when the findings are settled and the report is being written.

The core mental model

Sort every disclosure into one of two buckets:

• Necessary-minimum (keep): the endpoints the client calls, the request/response shapes it actually sends and receives, its own retry/backoff/debounce choices, input validation caps, and the status → UI behavior mapping. A public client cannot hide these; pretending otherwise is security theater.
• Over-disclosure (fix): anything describing what the server does with a request, server-side limits/quotas the client doesn't strictly need, endpoints the shipped client never calls, error branches that mirror internal server logic, anti-abuse mechanics, infra/tech-stack identifiers, private repo/paths, and test scaffolding that encodes backend behavior.