NS Security Audit

Find exploitable trust-boundary failures across the Expo app, TanStack web app, and Convex backend, repair confirmed issues at their shared enforcement point, and prove the repair without creating new operational risk.

Choose the Operating Mode

Infer the narrowest mode from the request:

• audit: inspect and report only; do not edit.
• audit and fix, remediate, or find and fix: inspect, implement the smallest complete repairs, add regression coverage, and verify locally.
• production verification: perform only explicitly authorized, low-impact read-only checks. Source-repository access never implies authorization to attack a deployed app, API, or provider account.

Treat authorization to edit this repository as authorization for local source changes and local tests only. It does not authorize deploys, provider changes, production writes, account creation, emails, payments, builds, store uploads, or destructive actions.

Hard Safety Boundaries

These rules override convenience and test completeness:

1. Never create or use a temporary, disposable, catch-all, generated, or agent-controlled email inbox or alias. Never invoke an AgentMail skill, tool, API, or CLI.
2. Never register an account with a third party's, employee's, founder's, published, guessed, or unverified email address. Never test pre-hijacking or account enumeration against real people.
3. Never trigger repeated signup, OTP, invitation, verification, or password reset emails. Do not brute-force, spray credentials, fuzz at volume, load test, race production requests, replay live tokens, or attempt takeover.
4. Never complete a real or test payment against production, validate a live store receipt by mutation, issue refunds, activate entitlements, create provider resources, or cross a confirmation boundary without explicit approval for that exact mutation.
5. Never print, copy, commit, or report secrets, session cookies, refresh tokens, store receipts, API keys, full personal data, or raw production rows. Redact evidence.
6. Never mutate or delete production data. Do not use destructive shell or Git commands. Preserve unrelated worktree changes.

For authenticated local verification, reuse the owner-provided account named by the repository verification rules. Follow ns-ios-verification for Simulator proof. If safe proof requires a forbidden action, report NOT VERIFIABLE and state the missing evidence.

Load the Required Context

1. Read AGENTS.md and every matching .agents/rules/* file before planning edits. Any Convex work also requires convex/_generated/ai/guidelines.md in full.
2. Read references/audit-controls.md for the mobile control matrix. Read references/evidence-and-reporting.md before assigning severity or writing the final report.
3. Inspect git status --short, package/runtime constraints, existing tests, and recent security-related changelog entries.
4. Run bash .agents/skills/ns-security-audit/scripts/security-audit-scan.sh . for candidate files. The scanner reports paths only and never confirms a vulnerability; inspect every candidate manually.

Audit Workflow

1. Map Trust Boundaries

Map Expo routes and native modules, TanStack public/auth/admin routes, Convex public/internal functions, Better Auth endpoints and token storage, Apple IAP, Stripe actions and webhooks, R2 uploads, deep/universal links, push tokens, background jobs, and client-visible DTOs. Identify the server-side authority for identity, role, entitlement, price, payment state, limits, and ownership.

2. Trace High-Risk Invariants End to End

Start with these chains before low-severity headers:

1. Apple/Google/Stripe proof -> provider verification and environment -> stable user -> server-priced purchase -> idempotent entitlement;
2. app session -> verified identity -> role -> object ownership -> minimal DTO;
3. public client config -> bundle boundary -> server-only Convex secrets;
4. deep/universal link -> trusted scheme/host -> validated route and callback;
5. upload request -> user ownership -> size/type/content validation -> safe object key -> isolated serving origin;
6. user URL or webhook target -> scheme/host/IP validation -> bounded fetch with redirect revalidation and timeout.

Do not trust client-provided user IDs, role, plan, product ID, price, paid state, receipt status, Stripe mode, redirect URL, MIME type, or object key. A native binary is an untrusted public client, even when distributed through a store.

3. Confirm Findings

Prefer source proof plus focused automated tests. A suspicious pattern is a CANDIDATE, not a finding. Confirm the reachable entrypoint, attacker control, missing server-side guard, impact, and absence of a compensating control.

Use a second user only in local or isolated test state with project-owned fixtures. Never create external accounts or inboxes to obtain proof. Record secure negative checks separately so they are not later reported as bugs.

4. Remediate at the Choke Point

When fixing is authorized:

• enforce auth, role, ownership, entitlement, and payment state in shared server/Convex builders or domain services;
• verify store receipts and Stripe events server-side, bind them to canonical products/users/environments, and make entitlement writes idempotent;
• keep secrets out of Expo public config, bundles, logs, DTOs, and persisted client state; store only the minimum session material in secure storage;
• validate deep-link origins and callback destinations at both native and server boundaries;
• move provider-only writes to internal functions and project minimal DTOs;
• use strict input schemas, explicit transition allowlists, and atomic quotas;
• fail closed with generic client errors and detailed server-only logs;
• add a regression test for the exploit path and a neighboring valid path.

Fix the exploit chain, not only the first observed endpoint. Search for sibling call sites that rely on the same broken assumption. Keep unrelated findings and unrelated worktree changes separate.

5. Verify Proportionally

Run the narrowest relevant checks first, then repository gates from .agents/rules/development-commands.md. For Convex changes, run the required codegen/check path. For web behavior, use dev-browser; for native behavior, follow ns-ios-verification and use xcrun simctl.

Do not start, stop, restart, deploy, build for stores, or change an existing runtime unless the request and repository rules authorize it. Never treat green local tests as production proof.

6. Update Project Records

After code or skill changes, update CHANGELOG.md. Do not commit, push, deploy, or open a pull request unless explicitly requested.

Output Contract

Lead with the security verdict and whether fixes were made. For each item give:

• severity and status: CONFIRMED, FIXED, CANDIDATE, NOT REPRODUCED, or NOT VERIFIABLE;
• affected boundary and concise impact;
• redacted evidence with file/line or test reference;
• root cause and complete remediation;
• verification level: static, automated, local runtime, staging, or production.

End with residual risks and skipped checks. Never include reusable exploit payloads, live identifiers, tokens, receipts, personal data, or instructions that would enable abuse of an external target.