The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted data from external social media profiles and messages.
Ingestion points: Social media profile bios, recent activity, and incoming messages/DMs from X and LinkedIn (mentioned in 'Platform Rules' and 'Workflow').
Boundary markers: None explicitly defined to separate untrusted profile data from agent instructions.
Capability inventory: The skill utilizes x-api, browser control for LinkedIn/X, and Apple Mail desktop automation (SKILL.md).
Sanitization: No explicit sanitization or filtering of external social media content is described before the LLM processes it for 'scoring' or 'drafting'.
[COMMAND_EXECUTION]: The skill requests the use of potentially sensitive execution environments to interact with user data.
Browser Control: Instructs the agent to use browser control for LinkedIn and X analysis and drafting when APIs are unavailable.
Desktop Automation: Utilizes desktop automation for Apple Mail (Mail.app) to create drafts, which requires interaction with the host operating system's UI and applications.
[DATA_EXFILTRATION]: The skill accesses and processes sensitive personal information related to the user's social network and communications.
Sensitive Data Access: Accesses private following/connection lists, message histories, and recent activity on professional and social platforms.
External Tool Usage: Passes enriched data to external tools like Exa, lead-intelligence, and social-graph-ranker for scoring and discovery.

connections-optimizer