The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes a dynamic instruction loading pattern where content retrieved from an external API is treated as 'authoritative'. It explicitly instructs the agent to prioritize this external data over its own internal safety and training data, which creates an indirect prompt injection surface.
Ingestion points: Content is retrieved via the get_skill tool call (SKILL.md).
Boundary markers: None identified; instructions do not specify delimiters for retrieved content.
Capability inventory: The agent can perform data exfiltration via the request_accountant_review tool.
Sanitization: No validation or sanitization of retrieved markdown is specified.
[DATA_EXFILTRATION]: The skill provides a mechanism to send 'working papers' to an external API. In an accounting context, these papers likely contain sensitive financial data or PII.
Evidence: request_accountant_review({ jurisdiction, scenario, working_paper }) triggers a network request to openaccountants.com.
[EXTERNAL_DOWNLOADS]: The skill is configured to fetch logic and data at runtime from the vendor's MCP server located at https://www.openaccountants.com/api/mcp.
Evidence: The skill description and the get_skill workflow detail fetching authoritative data from an external URL.

openaccountants-us-mo