The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data scraped from external websites, creating a surface for indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context through raw-post.json and raw-reviews.json files generated by scraping scripts.
Boundary markers: The LLM prompts in prompts/deconstruct-viral.md and prompts/review-to-ad.md do not utilize delimiters or specific instructions to ignore malicious commands embedded within the scraped content.
Capability inventory: The skill can write files to the local output/ directory and interact with external video generation tools via the Higgsfield MCP.
Sanitization: No sanitization or filtering is applied to the external content before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill executes local Node.js scripts to handle data retrieval.
Evidence: SKILL.md instructs the agent to run node skills/viral-replicator/scripts/fetch-ig-post.mjs <URL> and node skills/viral-replicator/scripts/scrape-g2-reviews.mjs <URL>.
Risk: User-provided URLs are passed as command-line arguments, which could facilitate command injection if the execution environment does not sanitize shell metacharacters.
[EXTERNAL_DOWNLOADS]: The skill automates data extraction from various third-party domains.
Evidence: The provided scripts use the playwright library to browse and fetch data from instagram.com, tiktok.com, youtube.com, g2.com, and trustpilot.com based on user input.

viral-replicator