The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external CRM systems and spreadsheets to generate business reports, creating an attack surface where malicious input could influence agent behavior.
Ingestion points: Data is pulled from various external sources including Salesforce, HubSpot, Pipedrive, Close, Supabase, Google Sheets, Notion, and local CSV files (Step 1).
Boundary markers: The skill lacks explicit boundary markers or instructions to the LLM to ignore potentially malicious instructions embedded within deal names, loss reasons, or other data fields during the analysis and reporting phases (Steps 2 and 3).
Capability inventory: The skill possesses capabilities to write markdown files to the local file system and transmit data to external services including Slack, Email (via agentmail), Notion, and Google Sheets (Step 4).
Sanitization: There is no evidence of sanitization or filtering logic applied to the external data before it is interpolated into the prompts for report generation.
[DATA_EXFILTRATION]: The skill is designed to transmit potentially sensitive sales pipeline data, including deal values, account details, and strategic recommendations, to external platforms such as Slack, Notion, and Email (Step 4). While these are intended features for business reporting, they represent a legitimate path for sensitive data to be sent outside the local environment.

pipeline-review