# Client Side Trust Boundary > Client-Side Trust Boundary - Skill: `annatchijova/client-side-trust-boundary` (Agent Skill) - Install (CLI): `npx skillmds@latest add annatchijova/client-side-trust-boundary` - Raw SKILL.md: https://api.skillmd.com/api/skills/annatchijova/client-side-trust-boundary/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: annatchijova (https://skillmd.com/u/annatchijova) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/annatchijova/client-side-trust-boundary --- # Client-Side Trust Boundary The line where your code stops running on hardware you control and starts running on the user's device is the most under-respected trust boundary in software. Past that line, everything is negotiable: the attacker owns the CPU, the memory, the storage, the network stack, and the debugger. They can root the phone, attach Frida and rewrite any function at runtime, bypass the certificate pin, dump the "encrypted" SQLite with the key sitting three files over, decompile the APK, and flip the `isPremium` boolean. **A security decision that is made only on the client has not been made.** It has been suggested to an adversary who is free to decline. Two failure modes, and both come from mistaking "the code we shipped" for "the code that runs": - **Trusting a client-side check.** The disabled button, the client-validated price, the feature gate, the input validation, the "jailbreak detected → refuse" guard — the attacker simply removes it. Worse, root/jailbreak detection and anti-tamper are themselves client-side code, so they are defeated by the same instrumentation they were meant to catch. A control that runs in the attacker's process protects nothing. - **Hiding a secret on the client.** The hardcoded API key, the "encrypted" local store whose key is derived from something also on the device, the obfuscated algorithm. A secret that ships to the client is a disclosed secret; obfuscation raises the cost of reading it from minutes to an afternoon, which is not a boundary. Composes with the library: - **validate-at-the-boundary** — the server-side corollary: every client input and every client decision is re-validated on trusted ground or it is unvalidated - **agent-trust-boundaries** — the client is an untrusted execution context whose output cannot carry authority - **authorization-surface-mapping** — the client can craft any request; the server's authorization is the only real gate (IDOR/BOLA is this failure) - **secret-lifecycle-discipline** — a secret embedded in a client is already leaked; treat it as compromised from ship - **credential-material-triage** — an extracted client key is captured material; triage what it actually unlocks - **honest-degradation** — a control that can only live client-side (offline, DRM) is defeatable; design for the loss, do not pretend it holds --- ## Step 0 — Draw the boundary and declare the far side hostile Locate, explicitly, where execution crosses onto the user's device. Everything past that point — app logic, local storage, in-memory values, outbound requests, TLS endpoints the app talks to — is attacker-influenced. This is Firstness done honestly: not "the app does X," but "the app *asks* the device to do X, and the device answers to the user." Write the boundary before you reason about any control, because every claim downstream depends on which side of it the control lives. --- ## Step 1 — Enumerate the client-side decisions and treat each as void List every security-relevant decision the client makes, and mark each one **not a control until it is re-made server-side**: - Authentication/authorization checks ("is this user allowed to see this screen") - Input validation ("the app won't let you submit a negative quantity") - Pricing/business logic ("total computed in the app") - Feature/entitlement gates ("premium unlocked locally") - Integrity/anti-tamper/root-jailbreak detection For each, the question is: *if the attacker deletes or flips this, what stops the abuse?* If the answer is "nothing, the server trusts the client's word," that is the finding. The server must independently recompute the price, re-check the entitlement, re-authorize the action — because the request it receives may have been crafted by hand with no app involved at all. --- ## Step 2 — Treat every shipped secret as disclosed Anything the client must *have* to function, the attacker can *extract*: - Hardcoded API keys, tokens, signing keys → disclosed; triage their blast radius (`credential-material-triage`) and prefer a design where the client holds nothing of value (server-side proxying, per-user short-lived tokens minted after real auth). - "Encrypted" local storage whose key is on the same device → the key travels with the ciphertext; this is encoding, not protection. - Obfuscation → a cost multiplier, not a boundary. State honestly that it deters casual and scaled attackers and does nothing against a targeted one (base rate: a determined reverser with Frida/decompilers reads it routinely). --- ## Step 3 — Separate what the client CAN do from what the server LETS it do The client can issue *any* request — any endpoint, any object id, any parameter, regardless of what the UI exposes. Capability is unbounded; the only real gate is server-side authority. Every "the app only ever requests its own data" assumption is an IDOR/BOLA waiting to happen. Route authorization through `authorization-surface-mapping`: the server must check, on every request, that *this* principal may perform *this* action on *this* object — never inferring permission from the fact that a well-behaved client would not have asked. --- ## Step 4 — Rate anti-tamper honestly: speed bumps, not walls Certificate pinning, root/jailbreak detection, integrity attestation, obfuscation, and anti-debugging are worth having — they raise attacker cost, cut down casual and automated abuse, and are appropriate defense-in-depth. But they are speed bumps, and naming them as boundaries is where teams get breached. State for each what it actually buys (deters scaled/casual attacks, adds reverse-engineering cost) and what it does not (stops a targeted attacker instrumenting the process). Never let a security property *rest* on one. Play integrity / hardware attestation is the partial exception — it anchors below the app — and belongs to `root-of-trust-reasoning`, not to app code. --- ## Step 5 — When a control can only live client-side, design for its defeat Some requirements are inherently client-side: offline-first enforcement, DRM, rate-limiting a fully local action. You cannot make these unbreakable. Honest degradation means: assume the control is defeated, and bound the damage — server-side reconciliation that detects the tampered state after the fact, limits on what a compromised client can affect, and monitoring that treats client-reported state as a claim, not a fact. Say in the design "this is enforced only on the client and is therefore defeatable; here is what limits the blast radius," instead of shipping a client check and calling the requirement met. --- ## The one-line test If the only thing standing between an attacker and the abuse runs on the attacker's own device, then nothing stands between them and the abuse. Move the check to the server, or write down plainly that the requirement is unenforced and design for the loss.