The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from social media bios and video metadata to 'draft outreach in the creator's language'. An attacker could embed malicious instructions in their social media bio to manipulate the agent's drafting or negotiation behavior.
Ingestion points: Creator bios, video titles, and link-in-bio pages scraped from TikTok and other platforms.
Boundary markers: The instructions do not define clear delimiters or 'ignore embedded instructions' warnings for the processing of scraped text.
Capability inventory: Drafting emails, negotiating rates, and executing SQLite database operations.
Sanitization: No sanitization is specified for external content before it is used in LLM prompts.
[DYNAMIC_EXECUTION]: The skill implements a local CRM using SQLite and instructs the agent to 'read and write it in plain SQL'. Constructing SQL statements from scraped external data without explicit sanitization or parameterization poses a risk of SQL injection.
[COMMAND_EXECUTION]: The workflow relies on executing shell commands and external utilities including yt-dlp for video metadata extraction and dig for verifying MX records during email validation.

influencer-outreach