← MCP server and skill grades

skills

samber/developer-relations-skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
c3a1b6912791
Files read
554

This project gives AI agents guides and helper scripts for developer relations tasks such as writing documentation, managing open-source repositories, and planning community content. The biggest risk is that a malicious sitemap could trick its crawler script into reading local files or reaching private networks, while unchecked inputs or unfamiliar README instructions could alter authenticated GitHub requests or run untrusted commands.

What it can reach

Files
Reads local code repositories, documentation, dependency lists, and exported logs, and writes drafted plans, temporary files, optional agent instructions, and project notes in devrel-context.md.
Network
Crawls documentation websites and sitemaps, queries GitHub APIs and api.deps.dev, and fetches public web pages like conference listings and platform policies.
Shell commands
Runs bundled Bash and Python scripts, command-line tools like git, curl, and dig, and setup or build commands from a project's README during audits.
Secrets
Reads the GITHUB_TOKEN environment variable when set and sends it to api.github.com for authenticated GitHub requests.

Findings

2 medium3 low
  1. MediumSitemap crawler fetches arbitrary schemes and cross-origin or internal URLs

    skills/docs-seo/scripts/docs-seo-crawl.py:42

    read_sitemap() extracts every <loc> entry from a remote sitemap XML file without checking its URL scheme or matching its hostname to the target site, and passes each entry to urllib.request.urlopen(). Because Python's urlopen() supports file:// and ftp:// in addition to http(s):// and follows cross-host redirects, crawling an untrusted or compromised external sitemap can read local files or send requests to localhost and internal network services, exposing extracted titles, headings, and descriptions in the output read by the agent.

    Fix: Validate that all crawled URLs and redirect targets use http or https, match the target site's hostname, and do not resolve to loopback, link-local, or private IP addresses.

  2. MediumGraphQL and URL query injection in GitHub profile audit script

    skills/github-profile-optimization/scripts/profile-audit.py:60

    main() extracts the target account name solely by splitting sys.argv[1] on '/' without validating GitHub username syntax, then interpolates it into REST API paths and a raw GraphQL JSON string sent with Authorization: Bearer $GITHUB_TOKEN. If prompt injection tricks the agent into auditing a crafted target such as 'github?x=...', the '?' turns the payload into an ignored query parameter on the initial REST checks while injecting arbitrary JSON and GraphQL queries or mutations into the authenticated /graphql request.

    Fix: Validate the login argument against '^[A-Za-z0-9](?:[A-Za-z0-9-]{0,37}[A-Za-z0-9])?$' and construct the GraphQL request body with json.dumps() and GraphQL variables instead of string formatting.

  3. LowGit option injection via unvalidated --from and --to arguments

    skills/changelog-writing/scripts/collect-release-material.sh:50

    collect-release-material.sh passes the --from and --to values directly to git log and git diff without rejecting leading dashes or adding --end-of-options. If prompt injection in a ticket or pull request tricks the agent into passing an argument starting with '--output=/path/to/file', git log writes its output to that path and overwrites existing files.

    Fix: Reject --from and --to arguments that start with '-' and validate both revisions with 'git rev-parse --verify --end-of-options' before running git log or git diff.

  4. LowPredictable temporary file path in /tmp

    skills/oss-contributor-onboarding/scripts/contributor-path-audit.sh:98

    contributor-path-audit.sh redirects git author statistics to a predictable PID-based path (/tmp/.cpa_authors.$$) instead of using mktemp. On multi-user systems without symlink protection in /tmp, another local user could pre-create a symlink at that path to overwrite files owned by the user running the script.

    Fix: Use mktemp to create a secure temporary file with a trap to remove it on exit, or pipe git log output directly into awk without writing to /tmp.

  5. LowCold-run checklists instruct the agent to execute README commands without container isolation

    skills/oss-launch/references/readiness-checklist.md:68

    When checking launch readiness (and similarly during README audits in skills/readme-optimization/references/audit-scorecard.md), the instructions tell the AI agent to literally run the install and example commands written in README.md inside a fresh local directory. A fresh folder on your computer still exposes your files, environment variables, and saved credentials. While safe on your own trusted code, asking the agent to audit a cloned third-party repository or an unreviewed external pull request that modified README.md could lead it to run untrusted shell commands on your machine if shell execution is approved.

    Fix: Require explicit user confirmation before executing any commands taken from repository documentation, and require running cold-run checks inside an isolated, unprivileged container with no mounted host credentials or network access where unneeded.