The Agent Skills Directory

REMOTE_CODE_EXECUTION (LOW): The skill documentation (SKILL.md and gcloud-platform.md) recommends installing the Google Cloud SDK using 'curl https://sdk.cloud.google.com | bash'. Piped remote execution is a high-risk pattern; however, since the source is a verified trusted organization (Google), the severity is downgraded to LOW per the [TRUST-SCOPE-RULE].\n- EXTERNAL_DOWNLOADS (LOW): The skill requires and recommends various official tools including the Wrangler CLI (npm install -g wrangler), Docker, and the Google Cloud CLI. These are standard industry dependencies from reputable sources.\n- COMMAND_EXECUTION (SAFE): The 'scripts/cloudflare_deploy.py' script executes the Wrangler CLI via 'subprocess.run'. It uses a safe execution pattern by passing arguments as a list and restricting the 'env' parameter to a set of predefined choices, effectively preventing shell injection.\n- PROMPT_INJECTION (LOW): Examples in 'references/browser-rendering.md' and 'references/cloudflare-workers-advanced.md' demonstrate scraping external websites and passing the content to an LLM (env.AI.run). This represents an indirect prompt injection surface. Evidence: Untrusted data enters via 'page.content()'; no boundary markers or explicit sanitization are shown in the code snippets; the skill has the capability to execute AI inference and network requests.

devops