The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes untrusted data from external RSS feeds which are then summarized by the AI.
Ingestion points: The scripts/fetch-feeds.js script fetches content from URLs defined in feeds.json.
Boundary markers: Absent. The synthesis prompt in SKILL.md (Step 4) interpolates article content directly into the instruction block without robust delimiters to distinguish source data from instructions.
Capability inventory: The agent can execute shell commands via Node.js scripts and has file system access to read/write temporary JSON data.
Sanitization: scripts/fetch-feeds.js performs basic HTML stripping for excerpts, but no sanitization or escaping is performed on the text to prevent embedded prompt instructions from influencing the LLM.
[COMMAND_EXECUTION]: The skill executes local JavaScript files using Node.js to perform core logic. Notable findings include:
Hardcoded paths: SKILL.md uses absolute paths (e.g., /Users/ksd/Desktop/Varnan_skills/newsletter-digest/) for script execution and file checks, which may cause errors or path traversal risks in different environments.
Verification logic: The agent is instructed to run node -e to manually parse and verify the JSON output of the fetch script.
[EXTERNAL_DOWNLOADS]: The skill programmatically fetches data from external RSS/Atom feed URLs specified in the feeds.json configuration file.

newsletter-digest