The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions (SKILL.md, Step 2) and README.md recommend and automate the installation of a third-party Node.js package defuddle using npm install -g. This package originates from an external GitHub repository (github.com/kepano/defuddle) which is not within the defined trusted vendor scope.
[REMOTE_CODE_EXECUTION]: The skill facilitates the installation and execution of code from an external repository. While it includes a user confirmation step ("Install it now... (yes/no)"), the capability to execute arbitrary logic from a remote package represents a significant attack surface.
[COMMAND_EXECUTION]: The skill executes shell commands using the defuddle CLI. Instructions like defuddle parse <url> --md interpolate user-provided URLs directly into shell commands, which could lead to command injection if the underlying tool or the agent's execution environment does not perform strict sanitization of the URL string.
[PRIVILEGE_ESCALATION]: The instruction to use npm install -g targets a global installation, which typically requires elevated system permissions (e.g., sudo or Administrator) on many environments.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process untrusted external data from URLs provided by the user.
Ingestion points: Web content fetched via defuddle or the WebFetch fallback (SKILL.md, Step 3a and 3b).
Boundary markers: The skill uses basic Markdown headers, source attribution strings, and horizontal rules (---) to separate extracted content (SKILL.md, Step 5).
Capability inventory: The skill possesses capabilities for package installation (npm install), arbitrary shell command execution (defuddle), and local file writing (-o flag).
Sanitization: There is no evidence of sanitization or filtering to prevent the agent from following instructions embedded within the fetched Markdown content.
[REMOTE_CODE_EXECUTION]: Automated security scans identified https://www.trulens.org/, which is included as an example URL in README.md and evals/evals.json, as a potentially malicious domain (URL:Blacklist).

webpage-reader