CanopiiCanopiiAll serversEnterprise →

android-agent MCP Server Security Report

io.github.ghost-in-the-droid/android-agent

Give any LLM agent a real Android or iPhone as its body. 62 MCP tools for mobile automation.

Security Trust Score
Grade F · Serious issues
Tier 1 · Public sourceHigh confidence · 92%latest v1.3.2 Repository
Declared MCP capabilities
ResourcesPrompts

This server scored F. Run it behind runtime policy enforcement so one bad tool call can't become an incident.

2 issues capping this score
  • criticalNo command-injection sinks — 2 occurrences
  • mediumNo over-broad / destructive tools — 3 over-broad tool(s)
Reputation244stars26forks0open issuestodaylast commit<1yage

Security controlslatest scored version v1.3.2

Each control is evaluated deterministically with evidence. The score is earned from passing controls; a failed guard caps it.

Model–MCP Runtime Guardrails

  • Fail
    User-in-the-Loop / Approval Scopeguard3 over-broad tool(s)

    No over-broad or destructive tools (arbitrary shell, bulk-delete) that warrant human approval.

    swipelist_skillscreate_skill

    Fix: Scope tools narrowly; avoid arbitrary command execution and destructive defaults.

  • Warn
    Strict JSON Schema Enforcementschemas not strict

    Tool inputs are constrained (additionalProperties:false), so unexpected arguments can't be smuggled in.

    Fix: Set additionalProperties:false and require explicit parameters on every tool.

  • Pass
    Indirect Prompt Injection (IPI) Defensesguardno injection markers

    Tool/prompt/resource text is free of hidden instructions that could hijack the agent.

  • N/A
    Tool Definition Integrityguardno prior version to diff

    Every consecutive version pair is diffed for new injection markers or destructive scope — a risky diff anywhere in history is a rug-pull (fail, durable); benign description drift warns.

Application Security Checks

  • Fail
    No command-injection sinksguard2 occurrences

    Untrusted tool input reaching a shell yields remote code execution.

    scripts/record_demo.py:432scripts/record_demo.py:470

    Fix: Never pass tool arguments to a shell; use argv arrays with shell disabled.

  • Warn
    No path traversalguard1 occurrence

    Naive path checks let tools read/write outside intended directories (EscapeRoute-class).

    site/src/lib/showcase.ts:53

    Fix: Resolve to a canonical path and verify containment; reject ../ and symlinks.

  • Warn
    No SSRF sinksguard35 occurrences

    Fetching tool-supplied URLs can pivot into internal networks and metadata services.

    frontend/src/composables/useApi.ts:3gitd/benchmarks/runner.py:230gitd/bots/common/adb.py:371gitd/bots/common/ios.py:786

    Fix: Allow-list destinations; reject arbitrary/loopback/link-local URLs.

  • Warn
    No known-vulnerable dependencies5 of 472 runtime deps vulnerable; worst high

    Runtime dependencies (parsed from the lockfile) are scanned against OSV.dev for published CVEs. Advisory: flagged dependencies lower the score but don't hard-cap it, since transitive reachability is unproven.

    [email protected] — GHSA-qx2v-qp2m-jg93[email protected] — GHSA-2pvr-wf23-7pc7[email protected] — GHSA-8hv8-536x-4wqp[email protected] — GHSA-j687-52p2-xcff

    Fix: Upgrade the flagged dependencies past their fixed versions.

  • Warn
    Established maintainerrecently published

    Brand-new / single anonymous maintainers raise takeover and malware risk.

    Fix: Publish under an established account/org; add multiple maintainers.

  • Pass
    No dynamic code executionguardno sinks found

    eval()/exec()/Function() on tool-derived strings allows arbitrary code execution.

  • Pass
    No unsafe deserializationguardno sinks found

    pickle/yaml.load/etc. on untrusted data can execute code.

  • Pass
    No committed secretsguardno secrets found

    Hardcoded keys/tokens in published source are live credentials an attacker can use.

  • Pass
    Credentials sourced from environmentreads credentials from environment

    Reading secrets from env/secret stores avoids hardcoding them.

  • Pass
    Dependencies pinned (lockfile)lockfile present

    A lockfile makes installs reproducible and resistant to silent dependency swaps.

  • Pass
    Package name not typosquattingguarddistinct package name

    Names mimicking popular packages are a common malware delivery vector.

  • Pass
    Actively maintainedrecent commits

    Unmaintained servers don't receive security fixes.

  • Pass
    Repository not archivedguardactive

    Archived repositories will never be patched.

  • Pass
    Declares a licenseMIT

    A clear license is required for legal enterprise use.

  • Pass
    Has a security policySECURITY.md present

    A SECURITY.md gives a private path to report vulnerabilities.

  • Pass
    Adoption & popularityestablished adoption

    A small, capped nudge from stars/downloads — widely-used servers get more eyes on bugs. It can never offset a real security failure.

  • Not checked
    No install/post-install scriptsguardnot determinable from PyPI metadata

    install hooks run arbitrary code on every consumer at install time.

  • Not checked
    Published with provenancenot determinable

    Build provenance attests the artifact was built from the claimed source by CI.

  • Not checked
    Signed releasesnot evaluated

    Signed releases let consumers verify artifacts weren't tampered with.

Transport & Trust Model

  • Warn
    Execution Sandboxingruns natively (no container image)

    A container/sandbox image limits blast radius; a server that runs natively has full host access.

    Fix: Ship a Dockerfile/Containerfile (or document a sandboxed run) so the server runs isolated.

  • Warn
    Network Exposurebinds 0.0.0.0

    Binding 0.0.0.0 or exposing debug inspectors widens the attack surface.

    Fix: Bind to 127.0.0.1 by default; never ship open debug endpoints.

  • N/A
    Transport Encryption (TLS)guardno remote endpoints

    Plaintext HTTP exposes traffic and bearer tokens to interception.

  • N/A
    IAM / Authentication Scopingno remote endpoints

    OAuth 2.1 / Protected Resource Metadata gates who can invoke tools.

  • N/A
    Live Endpoint Reachablenot dynamically scanned

    A dynamic scan connected to the declared remote endpoint and it responded — verified live, not a dead URL.

  • N/A
    Authentication Enforced (live)not dynamically scanned

    If the server declares auth is required, it must actually reject anonymous clients. Serving tools to unauthenticated callers is a real exposure.

Embed this score

Add the live badge to your README — it updates automatically on every rescan.

Canopii Trust Score badge preview
markdown[![Canopii Trust Score](https://index.canopii.dev/api/badge/io.github.ghost-in-the-droid/android-agent)](https://index.canopii.dev/server/io.github.ghost-in-the-droid/android-agent)
html<a href="https://index.canopii.dev/server/io.github.ghost-in-the-droid/android-agent"><img src="https://index.canopii.dev/api/badge/io.github.ghost-in-the-droid/android-agent" alt="Canopii Trust Score" /></a>

Versions

VersionScoreStatus
v1.3.2latest14Fscored