Oathe Security Badge

oathe / scan / reports / caution

Is TrialAndErrorAI/appstore-connect-mcp safe?

Use with caution. TrialAndErrorAI/appstore-connect-mcp is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the appstore-connect-mcp skill by TrialAndErrorAI a trust score of 72/100 with 7 findings (2 high). Report updated

https://github.com/TrialAndErrorAI/appstore-connect-mcp

72
CAUTION

Is TrialAndErrorAI/appstore-connect-mcp safe to install?

appstore-connect-mcp is a legitimate, purpose-built MCP server for Apple App Store Connect API access with no prompt injection content in SKILL.md, no malicious install scripts, and clean clone-time behavior. The primary risks are architectural: the execute() tool's Node.js vm sandbox provides weak isolation for LLM-generated code that runs with an authenticated Apple API client attached, and an Admin-scoped API key enables write operations across pricing, metadata, and beta distribution. Operators should understand they are granting an AI agent persistent, broad access to their App Store Connect account including sensitive financial data and app management capabilities.

What security issues were found in TrialAndErrorAI/appstore-connect-mcp?

Category Scores

Prompt Injection 95/100 · 30%
Data Exfiltration 60/100 · 25%
Code Execution 58/100 · 20%
Clone Behavior 92/100 · 10%
Canary Integrity 100/100 · 10%
Behavioral Reasoning 58/100 · 5%

Findings (7)

HIGH execute() tool runs LLM-generated JS in a weak vm sandbox -30 ▶

The core value proposition of this MCP is that the LLM writes JavaScript which is executed server-side via Node.js vm.runInNewContext(). Node.js vm is explicitly documented as insufficient for isolating untrusted code — it shares the same V8 heap and prototype chain, enabling escape via proto manipulation and similar techniques. The sandbox blocks fetch, require, and process, which mitigates the most obvious vectors, but does not provide the isolation equivalent to a separate process or Isolate. Any exploit that escapes the vm context gains access to the Node.js runtime with the full credentials loaded from environment variables.

Evidence: src/executor/sandbox.ts uses vm.createContext + vm.runInNewContext; RFC-002 acknowledges 'Phase 2: Consider isolated-vm for stricter isolation if publishing to npm' — confirming the current implementation is known to be weaker

HIGH execute() exposes authenticated Apple ASC client capable of reading all financial and user data -25 ▶

The api global injected into every execute() sandbox call is an authenticated HTTP client wrapping the full App Store Connect API — 923 endpoints covering revenue, subscription events, paying users, TestFlight tester emails, crash data, and more. Code running in the sandbox can read this data and return it to the LLM context, where it may be leaked via LLM output, stored in conversation history accessible to other skills, or used as inputs to further API calls. This is intentional functionality, but the data sensitivity warrants explicit operator awareness.

Evidence: API-COVERAGE.md documents access to salesReports, financeReports, customerReviews, betaTesters, subscriptions — all readable via api.request() in execute(). SETUP.md recommends Admin-scoped API key.

MEDIUM Admin API key enables destructive write operations via execute() -20 ▶

SETUP.md recommends generating an API key with 'Admin' role for 'full access'. The execute() tool can call POST/PATCH/DELETE endpoints on the App Store Connect API, enabling modifications to app pricing across territories, app metadata (title, keywords, description), screenshot sets, phased release controls, IAP configurations, subscription pricing, and beta tester lists. A sophisticated prompt injection from another source could instruct the LLM to issue destructive changes through this skill.

Evidence: RFC-002 explicitly discusses enabling write scope: 'POST/PATCH endpoints: Code mode enables WRITE operations. Do we ship those in v2.0 or add in v2.1?' — suggesting write operations are in scope; API-COVERAGE.md lists WRITE endpoints for appInfoLocalizations, appStoreVersions, subscriptions, pricing.

MEDIUM vm sandbox escape could grant access to p8 private key in environment -15 ▶

The Apple ASC JWT is generated from a P8 private key loaded from the filesystem path specified in APP_STORE_P8_PATH environment variable. If the vm sandbox is escaped (a known class of vulnerability against Node.js vm), attacker code could read process.env.APP_STORE_P8_PATH, then use Node.js fs to read the P8 key file, exfiltrating the signing credential. The P8 key cannot be revoked from Apple's side without manual intervention and can be used to impersonate the developer indefinitely.

Evidence: src/auth/jwt-manager.ts loads P8 key via fs.readFileSync(config.p8Path) at startup; sandbox.ts sets process: undefined in context but this only shadows process in the vm context — a vm escape bypasses this.

LOW No install-time code execution — clean package lifecycle 0 ▶

package.json contains no preinstall or postinstall scripts. The prepare and prepublishOnly hooks only run 'npm run build' (tsc compilation + file copy + chmod +x). No git hooks, no .gitattributes filter drivers, no submodules. Installation is safe.

Evidence: package.json scripts section: prepare and prepublishOnly both run 'npm run build'; no preinstall/postinstall keys present

INFO SKILL.md is empty — no prompt injection surface 0 ▶

The SKILL.md file, the primary vector for prompt injection in MCP skills, contains no content. No instructions are injected into the agent's system prompt via this mechanism.

Evidence: SKILL.md content in evidence is empty string

INFO All canary files intact — no credential harvesting detected 0 ▶

Honeypot files (.env, id_rsa, .aws/credentials, .npmrc, .docker/config.json, GCP credentials) were accessed only by the audit framework's own sudo-based setup and teardown processes. The timing (pre-clone and post-install phases) and process context (sudo) confirm these were not caused by the skill under test.

Evidence: Canary integrity: ✅ All canary files intact; auditd PATH records at 1794735234 show sudo (ppid=1083, pid=1101) — the audit harness, not git clone or npm

Should I install TrialAndErrorAI/appstore-connect-mcp?

Oathe's verdict for TrialAndErrorAI/appstore-connect-mcp is CAUTION with a trust score of 72/100. Recommendation: Install with caution.