The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to access and process highly sensitive personal data, including the contents of the user's Gmail inbox/sent folders and Google Calendar events. While this is necessary for its stated purpose as a CRM, the broad scope of data access increases the risk profile.
Evidence: The configuration requires GMAIL_OAUTH_CREDENTIALS and GOOGLE_CALENDAR_CREDENTIALS for full operation.
Evidence: The skill transmits processed data summaries and contact information to external well-known services including Telegram, Apollo.io, and Brave Search.
Note: These behaviors are transparently documented as intended functionality.
[PROMPT_INJECTION]: The skill is inherently vulnerable to indirect prompt injection due to its core function of processing untrusted external communications.
Ingestion points: The agent-data-collector.md specification details the extraction of data from Gmail (INBOX/SENT), Google Calendar attendees/descriptions, and meeting notes files.
Boundary markers: The LLM prompts used for data extraction (e.g., in agent-action-extractor.md and agent-relationship.md) do not implement explicit boundary markers or delimiters to protect against instruction override within the ingested data.
Capability inventory: The skill has the capability to write to local memory files, perform network operations to multiple APIs, and generate message drafts for the user to send.
Sanitization: The agent-data-collector.md mentions stripping HTML tags from snippets, but no semantic sanitization or instruction filtering is applied to the raw content before LLM processing.

self-evolving-crm