← MCP server and skill grades

skills

samber/cc-skills-golang/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
d68d80aec298
Files read
299

This project is a collection of 33 AI coding agent skills that guide Go development, testing, profiling, refactoring, security auditing, and CI/CD setup. Its biggest risk is that the skill definitions pre-approve broad file-writing, shell (go, git, gh, curl), and web-fetching tools that activate automatically on Go files, meaning prompt injection from untrusted code or web pages could run code or exfiltrate data without confirmation prompts.

What it can reach

Files
Reads, edits, and creates files in the project workspace, including Go source code, GitHub Actions workflows, and agent configuration files such as CLAUDE.md and AGENTS.md.
Shell commands
Executes Go toolchain and developer commands (go, golangci-lint, gopls, dlv, govulncheck, benchstat, swag, wire, protoc, goreleaser, godig), git commands, GitHub CLI (gh) commands, perf, and curl.
Network
Fetches web pages and search results via WebFetch and WebSearch, sends HTTP requests via curl, queries pkg.go.dev and vuln.go.dev (or an optional hosted server at godig.samber.dev), and calls Context7 documentation MCP tools.

Findings

1 high4 low1 info
  1. HighOver-broad allowed-tools auto-approvals on Go skills, including read-only and path-triggered skills

    skills/golang-benchmark/SKILL.md:21

    Nearly all 33 skills declare broad allowed-tools wildcards in their frontmatter that pre-approve file writes (Edit, Write), code execution (Bash(go:*)), and git operations (Bash(git:*)) without asking the user for confirmation, and 26 skills set paths to **/*.go so they activate automatically whenever the agent opens a Go file (golang-graphql even sets user-invocable: false while auto-activating on Go files). Several skills also pre-approve unrestricted Bash(curl:*) alongside WebFetch (golang-benchmark, golang-performance, golang-graphql) or unrestricted Bash(gh:*) for all GitHub CLI operations (golang-continuous-integration, golang-refactoring), and skills explicitly described as read-only or advisory (golang-pkg-go-dev, golang-stay-updated, golang-popular-libraries) still grant Edit, Write, Bash(go:*), and Bash(git:*). If the agent reads prompt-injected text in an untrusted repository, issue, or fetched web page, it can write and run arbitrary Go code, upload local files with curl, or perform GitHub account actions without confirmation prompts.

    Fix: Remove Edit, Write, Bash(go:*), and Bash(git:*) from read-only skills, drop Bash(curl:*) or restrict it to localhost pprof URLs where needed, scope Bash(gh:*) to specific subcommands instead of a wildcard, and avoid pairing write/exec auto-approvals with automatic **/*.go path activation.

  2. LowLocal HTTP servers for MCP, pprof, and pkgsite bind to all network interfaces without authentication

    skills/golang-pkg-go-dev/SKILL.md:66

    The setup instructions start the shared godig HTTP MCP server on :8080 with no authentication even though the client command on the very next line connects to localhost:8080. Likewise, golang-benchmark/references/pprof.md (go tool pprof -http=:8080) and golang-documentation/references/library.md (go tool pkgsite -http=:6060) omit the loopback address before the port. In Go, :port listens on 0.0.0.0 and [::] across all network interfaces, exposing local source code, memory/CPU profiles, or the MCP server to anyone on the same local network.

    Fix: Bind local servers explicitly to loopback (127.0.0.1:8080 or localhost:8080) instead of :8080 or :6060.

  3. LowGenerated Claude CI review workflow ingests untrusted PR files into the agent prompt

    skills/golang-continuous-integration/assets/claude-code-review.yml:98

    When golang-continuous-integration sets up AI PR review from this workflow asset, each job checks out the untrusted pull request branch and directly includes @./CLAUDE.md, @./CONTRIBUTING.md, and @./docs/project-summary.md in the review prompt, while the discuss job also triggers on review comments from any non-bot GitHub user. Because the workflow runs with repository-wide pull-requests: write and actions: read permissions, ANTHROPIC_API_KEY, Bash(gh pr comment:*), and external Context7 MCP tools, a malicious PR author or commenter can plant prompt-injection text in CLAUDE.md or a comment to post arbitrary PR comments or leak data from workflow run logs.

    Fix: Do not load CLAUDE.md, CONTRIBUTING.md, or docs from the untrusted PR head checkout (read trusted base-branch files instead), restrict workflow triggers to trusted contributors, and scope permissions per job.

  4. LowUnpinned third-party tool and skill installations via npx and @latest

    skills/golang-continuous-integration/SKILL.md:258

    Both the Copilot setup command and the generated claude-code-review.yml CI workflow execute npx skills add without pinning the skills npm package version and pull the mutable default branch of github.com/samber/cc-skills-golang with -y (including inside CI jobs that hold ANTHROPIC_API_KEY and pull-requests: write). In addition, 12 SKILL.md files declare unpinned @latest Go binaries (such as godig@latest, gotests@latest, swag@latest, wire@latest, and dlv@latest) and the unversioned skills npm package in their install metadata and dependency instructions. If an upstream package or repository is compromised, unpinned installs execute the attacker's code automatically during setup or CI runs.

    Fix: Pin the skills npm package to an exact version, pin GitHub repository URLs in CI workflows to a verified commit SHA or release tag, and pin Go tool install instructions to specific release versions.

  5. LowProject-layout skill modifies agent configuration files without user confirmation

    skills/golang-project-layout/SKILL.md:119

    The Initialization Checklist in golang-project-layout instructs the agent to write an always-load directive for golang-how-to into CLAUDE.md, AGENTS.md, GEMINI.md, .github/copilot-instructions.md, or .cursor/rules/golang-skills.mdc (with alwaysApply: true) and explicitly states that no user confirmation is needed. A developer who only asked to scaffold a Go project layout gets persistent instructions added to their agent config that automatically load skills on future tasks.

    Fix: Ask the user for confirmation before creating or modifying CLAUDE.md, AGENTS.md, Cursor rules, or Copilot instruction files during project initialization.

  6. InfoOptional third-party hosted MCP endpoint receives package queries and supplies tool metadata

    skills/golang-pkg-go-dev/SKILL.md:73

    As an optional fallback when the godig CLI is not installed locally, golang-pkg-go-dev provides a command to register a public MCP server at https://godig.samber.dev/mcp. Users who opt into the hosted server send their package and module lookup paths (which could include private internal module names if queried by the agent) to a third-party host and rely on that remote server for live MCP tool responses.

    Fix: Use the local godig CLI or stdio MCP server (godig mcp) rather than the public hosted instance, especially in repositories with private Go module paths.