The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted external sources, including CRM records and general web research, which could contain malicious instructions designed to influence the agent's scoring or output.
Ingestion points: CRM search results (account records, deal history, engagement history) and web-based company research (Step 1a and 1b in SKILL.md).
Boundary markers: The skill uses template placeholders like {Client Profile: X} to load configuration, but lacks explicit instructions or boundary markers to treat ingested CRM or web data as non-executable content.
Capability inventory: The skill has access to tool-based web/CRM search and the ability to generate files (Markdown, HTML, PDF).
Sanitization: There is no mention of sanitizing or filtering external content before it is processed for scoring.
[DYNAMIC_EXECUTION]: The skill instructions suggest using Python and the reportlab library to generate PDF qualification cards.
Evidence: Option C: PDF -- Python + reportlab under the "Artifact Generation" section of SKILL.md.
This implies the agent may generate and execute a Python script at runtime to create the document.
[DATA_EXFILTRATION]: The skill requires access to highly sensitive organizational data (CRM records, contacts, and pipeline deals) and subsequently performs web research and output generation.
Evidence: Core workflow steps 1a and 1b in SKILL.md involve pulling extensive CRM data and then researching the company on the web. While no malicious exfiltration target is identified, the combined capability to read sensitive data and output files/perform web requests constitutes a data exposure surface.

gtm-account-qualification