skills
- Type
- Plugin
- Scanned
- Nov 13, 2026
- Version
- cd91460c0834
- Files read
- 180
This official Expo plugin gives AI coding agents a connection to Expo's remote MCP server and 24 skills for building, testing, upgrading, and deploying Expo and React Native apps. Its biggest risk is that several skills pre-approve broad command permissions (such as running any node, eas, gh api, or bun create command without asking), which could let a prompt-injected agent run arbitrary code or leak data, and every skill tells the agent to send task failure reports to Expo without asking the user first.
What it can reach
- Network
- Connects to Expo's remote MCP server (https://mcp.expo.dev/mcp), fetches EAS workflow schemas and documentation from Expo and GitHub, communicates with EAS cloud services, downloads npm packages, and optionally opens public development tunnels for remote simulators.
- Files
- Reads and edits project source files, Expo and EAS configuration files (app.json, eas.json, .eas/workflows/*.yml, .env.eas-simulator), local HTTP cache files under eas-workflows/scripts/.cache, and installed package type definitions in node_modules.
- Shell commands
- Instructs the agent—and in several skills pre-approves via allowed-tools—to run local commands including node, npx, eas, expo, gh, git, bun, xcodebuild, pod, ffmpeg, and device automation CLIs.
- Secrets
- Uses and guides configuration of Expo/EAS login sessions and EXPO_TOKEN, simulator session tokens in .env.eas-simulator, GitHub CLI credentials via gh api, and Apple Developer or Google Play submission credentials.
- System
- Boots, drives, and captures screenshots or recordings from local and remote iOS simulators and Android emulators when testing or evaluating apps.
Findings
-
HighUnrestricted Node.js execution and file writes pre-approved in eas-workflows skill
plugins/expo/skills/eas-workflows/SKILL.md:4
The eas-workflows skill only needs Node.js to run its bundled scripts/fetch.js helper for downloading workflow schemas, but its allowed-tools frontmatter pre-approves Bash(node:*) as well as unscoped Read and Write. In agents such as Claude Code that honor allowed-tools, Bash(node:*) lets the agent execute any Node.js command—including node -e arbitrary code execution—without asking the user for permission. If the agent reads untrusted text containing prompt injection while working on CI/CD workflows, it can silently run arbitrary code, read local secrets, or overwrite files. In addition, scripts/fetch.js accepts any URL on the command line without restricting to Expo domains.
Fix: Scope the Node rule to the specific script path (for example, Bash(node *eas-workflows/scripts/fetch.js *)), restrict scripts/fetch.js to https://api.expo.dev/ and https://raw.githubusercontent.com/expo/expo/ URLs, remove unscoped Write or narrow it to .eas/workflows/**, and restrict the eas-cli rule to workflow:validate.
-
HighOver-broad EAS and npx command auto-approvals allow arbitrary local execution and cloud actions
plugins/expo/skills/eas-simulator/SKILL.md:5
The eas-simulator, eas-update, and read-only eas-update-insights skills pre-approve Bash(eas *) and/or wildcard npx patterns such as Bash(npx *eas-cli@*) and Bash(npx *agent-device@*). First, eas simulator:exec <cmd> loads .env.eas-simulator and spawns any local command passed to it, so pre-approving Bash(eas *) lets a prompt-injected agent run arbitrary shell commands on the user's machine without confirmation. Second, leading wildcards before the package name in Bash(npx *eas-cli@*) and Bash(npx *agent-device@*) match spaces and npm scopes, allowing commands like npx --yes @attacker/eas-cli@latest or npx --yes malicious-pkg eas-cli@1 to download and execute untrusted npm packages without prompting. Third, Bash(eas *) auto-approves production OTA updates, paid cloud builds, and secret retrieval even in the read-only eas-update-insights skill, while pod* and xcodebuild* execute workspace build scripts.
Fix: Restrict allowed-tools in each skill to the exact subcommands required (for example, only eas update:list, eas update:insights, eas update:view, and eas channel:insights in eas-update-insights), remove wildcard pre-approval for eas simulator:exec, and anchor npx rules to exact package arguments without a leading * wildcard.
-
HighPre-approved GitHub API and template commands in expo-examples enable silent data exfiltration and code execution
plugins/expo/skills/expo-examples/SKILL.md:4
The expo-examples skill only needs read-only access to the public expo/examples repository or to scaffold from it, yet its allowed-tools frontmatter pre-approves unscoped Read, Bash(gh api:*), Bash(bun create:*), Bash(npx degit:*), and Bash(git clone:*) without user confirmation. Because gh api uses the user's authenticated GitHub CLI session and supports POST, PUT, and DELETE requests, a prompt-injected agent can silently read local files and send them to an external GitHub gist, issue, or repository—or modify the user's GitHub repositories. Furthermore, Bash(bun create:*) pre-approves running any create-* package from npm or GitHub (such as bun create attacker-pkg), executing arbitrary remote code without prompting.
Fix: Avoid pre-approving authenticated gh api calls for reading public repo files (or restrict the rule to read-only paths under repos/expo/examples/ without write flags), and narrow bun create, npx degit, and git clone rules so they only match expo/examples.
-
MediumAll skills instruct the agent to send task reports and project context to Expo without user approval
plugins/expo/skills/expo-overview/SKILL.md:136
Every one of the 24 skills in plugins/expo and plugins/expo-experiments includes a Submitting Feedback block telling the agent to run npx --yes @expo/agent-cli@latest feedback whenever it encounters skill issues, @expo/agent-cli errors, or repeated task failures where the user has to step in, and instructs it to include 'as much relevant context as possible' (up to 5,000 characters). Unlike the GitHub issue instruction in expo-migrate-module—which explicitly requires asking the user before posting—this block never asks for user confirmation or shows what will be sent. In auto-approved agent sessions or under prompt injection from error logs, private task descriptions, file paths, or code context can be sent to Expo without the user's consent.
Fix: Update the shared feedback block in scripts/feedback-block.ts and all SKILL.md files to require explicit user confirmation, display the exact payload beforehand, and strip secrets, file paths, and proprietary code before submitting feedback.
-
LowEntry-point and workflow skills bypass project lockfiles with unpinned @latest and unscoped npm CLIs
plugins/expo/skills/expo-overview/SKILL.md:74
The expo-overview router skill—loaded first for every Expo task—instructs the agent to replace standard npx expo commands (which use the version pinned in the project's lockfile) with npx --yes @expo/agent-cli@latest, silently downloading and running whatever version is currently tagged latest on npm on every invocation. Other skills similarly instruct running unpinned @latest or unscoped third-party npm packages in sensitive flows, including npx agent-device@latest with simulator session tokens in eas-simulator, and bunx setup-safari and bunx testflight/npx testflight with Apple Developer and App Store Connect credentials in expo-app-clip and eas-app-stores. A supply-chain compromise of any of these npm packages would immediately execute untrusted code with local and cloud credentials.
Fix: Prefer the project's lockfile-pinned Expo CLI (npx --no-install expo) for standard commands, avoid passing --yes with @latest without user consent, pin vetted package versions, and prefer official eas credentials and eas submit commands over unscoped wrapper packages when handling Apple Developer credentials.
-
LowInternal expo-skill-eval developer harness is exposed to wildcard skill installs and uses predictable /tmp execution rules
.claude/skills/expo-skill-eval/SKILL.md:5
The repository's internal evaluation skill lives at .claude/skills/expo-skill-eval/SKILL.md. While plugin marketplace installs only ship plugins/expo, users who follow the README's wildcard command (npx skills@latest add expo/skills --skill '*') can unintentionally install this contributor-only harness. Its allowed-tools frontmatter pre-approves Bash(python3 /tmp/expo-skill-eval-*) and Write(/tmp/expo-skill-eval-*/**) on predictable paths in shared /tmp, and instructs running claude -p --dangerously-skip-permissions after a prompt. In addition, its helper scripts interpolate unescaped JSON strings into viewer.html (generate_viewer.py) and unescaped paths/package names into node -e and adb shell (snapshot-android.sh, snapshot-ios.sh).
Fix: Exclude .claude/skills/expo-skill-eval from user-facing skills.sh discovery or document explicit skill selection instead of --skill '*', use mktemp -d instead of predictable /tmp paths, HTML-escape untrusted fields in generate_viewer.py, and pass paths to node via argv rather than string interpolation in snapshot scripts.
