← MCP server and skill grades

skills

cloudflare/skills/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
cbf893b95422
Files read
232

This project is Cloudflare's official collection of agent skills that guide AI coding assistants in building, configuring, and deploying applications across Cloudflare Workers, storage, networking, and Zero Trust services. Its biggest risk is that the Cloudflare One skill instructs the agent to automatically read Access service-token secrets from local secret stores and send them to any host that returns a Cloudflare Access login page, which could leak credentials if the agent fetches an attacker-controlled URL.

What it can reach

Files
Reads and edits project source code, Cloudflare and framework configuration files, local .env files, and agent skill or instruction files such as .claude/skills, CLAUDE.md, and AGENTS.md.
Shell commands
Instructs the agent to run terminal commands including cf, wrangler, npm, npx, git, curl, python3, jq, gh, and bundled bash scripts under skills/turnstile-spin/scripts/.
Secrets
Reads, validates, and uploads Cloudflare API tokens, Turnstile widget secrets, Cloudflare Access service tokens, and Worker or R2 credentials from environment variables, .env files, and secret stores.
Network
Contacts Cloudflare APIs, Cloudflare documentation, GitHub repositories, npm registries, local dev servers, and user- or task-referenced web URLs.
Browser
Uses browser automation and performance-tracing MCP tools such as Chrome DevTools MCP to load pages, record traces, and inspect network requests.

Findings

1 high3 medium4 low
  1. HighCloudflare Access service-token secrets can be sent to untrusted hosts spoofing an Access login page

    skills/cloudflare-one/SKILL.md:30

    When an agent encounters a Cloudflare Access login page or authentication redirect while fetching a URL, this skill instructs it to silently read CF-Access-Client-Id and CF-Access-Client-Secret from the project's secret store or environment variables and resend them as headers on the original request. Although it warns against forwarding headers across cross-origin redirects, it does not require user approval or check an explicit hostname allowlist for the original URL, so prompt injection that tricks the agent into fetching an attacker-controlled URL returning a fake Access login page could exfiltrate the service token.

    Fix: Require explicit user confirmation of the target hostname, or verify via the Cloudflare API that the hostname belongs to an Access application in the user's account, before sending CF-Access-Client-Secret headers.

  2. MediumNext.js skill runs unpinned npx package and installs remote skills without asking

    skills/nextjs-on-cloudflare/SKILL.md:3

    Both the skill description and workflow tell the agent to automatically run npx skills add cloudflare/vinext whenever upstream vinext skills are missing, and then follow the downloaded SKILL.md instructions. This downloads and executes an unpinned npm package via npx and pulls unpinned skill files from GitHub at runtime without first asking the user for permission.

    Fix: Ask the user for confirmation before running npx or installing external skills, and pin both the CLI package version and the upstream repository commit or release tag.

  3. MediumCLI skill instructs global package installation and user-level agent config edits without confirmation

    skills/cf-cli/SKILL.md:45

    The setup section places npm install --global cf before checking whether cf is already installed and instructs the agent to add a preference rule to the user-level AGENTS.md or CLAUDE.md file. Unlike sibling skills that forbid unprompted global changes, this can cause an agent to install an unpinned global npm package and persistently alter agent behavior across all projects on the machine without the user asking.

    Fix: Check whether cf is installed first, ask the user before running a global npm install, pin a known version if needed, and ask explicit permission before editing user-level AGENTS.md or CLAUDE.md files.

  4. MediumBasin Pipeline templates enable public HTTP ingestion with authentication disabled

    skills/basin/references/pipelines/configuration.md:72

    Both the REST API and Terraform stream creation examples enable a public HTTP ingestion endpoint with authentication set to false. Because Basin streams are immutable once created, an agent copying these templates provisions an unauthenticated public endpoint that allows anyone who discovers the stream URL to inject forged records into the user's R2 Iceberg tables.

    Fix: Set authentication to true by default in the REST and Terraform stream examples and add an explicit warning about public access when authentication is disabled.

  5. LowFrameworks skill directs the agent to load unpinned skills from third-party GitHub branches

    skills/frameworks-on-cloudflare/SKILL.md:27

    When setting up frameworks such as TanStack Start, React Router, or Hono, this skill tells the agent to fetch and load SKILL.md files from the unpinned main branches of external GitHub repositories. Because loaded skills act as trusted instructions to the coding agent, an upstream branch compromise in any of those third-party repositories could inject malicious instructions at runtime.

    Fix: Pin external skill references to reviewed commit hashes or release tags and treat remote third-party markdown as untrusted reference material unless approved by the user.

  6. LowTurnstile persistence script clones unpinned GitHub HEAD and installs executable scripts

    skills/turnstile-spin/scripts/persist-skill.sh:67

    When saving the Turnstile Spin skill into a project, persist-skill.sh clones the unpinned default branch of https://github.com/cloudflare/skills.git over the network and marks the downloaded .sh scripts executable, even though a local copy of the skill bundle is already present on disk. Those scripts handle CLOUDFLARE_API_TOKEN and TURNSTILE_SECRET in later sessions.

    Fix: Copy the skill directory from the already-installed local bundle relative to persist-skill.sh instead of cloning an unpinned remote branch over the network.

  7. LowWeb performance skill recommends MCP setup using npx -y with an unpinned @latest tag

    skills/web-perf/SKILL.md:27

    The example MCP server configuration runs npx -y chrome-devtools-mcp@latest, which automatically downloads and executes whatever version is currently tagged latest on npm with local user privileges whenever the MCP server starts.

    Fix: Recommend pinning chrome-devtools-mcp to a specific vetted version instead of @latest.

  8. LowBasin CLI and curl examples pass secret keys and tokens in command-line arguments

    skills/basin/references/pipelines/configuration.md:52

    The Basin sink and stream setup examples pass --catalog-token, --secret-access-key, and curl Authorization headers directly in command-line arguments, exposing credentials in process listings and potentially in shell history or agent command logs if interpolated.

    Fix: Pass Authorization headers and secret-bearing payloads via standard input (such as curl --config - and --data-binary @-) and use environment variables or secret prompts where supported.