The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted sources, creating a surface for indirect injection attacks.
Ingestion points: The agent processes client messages in live dialogues, content from external websites via the scrape_url tool, and uploaded files (PDF, DOCX, XLSX) through the Vector Document and Table features.
Boundary markers: While the documentation suggests using a 'Restrictions' section in system prompts, there are no mandatory structural delimiters or 'ignore embedded instructions' markers defined for handling ingested data.
Capability inventory: The agent has access to powerful tools including webhooks, SQL query execution, account switching (active_user_id), and the ability to call other bots.
Sanitization: The provided instructions do not mention explicit sanitization, validation, or filtering of external content before it is interpolated into the agent's context or prompt templates.
[DYNAMIC_EXECUTION]: The platform uses runtime generation and execution of prompts and queries.
The Liquid template engine is used to dynamically render system instructions based on dialogue-scoped variables, allowing for conditional logic and variable interpolation at runtime.
The agent is tasked with writing and executing SQL queries against Tables (CSV, XLSX, or Google Sheets) to retrieve specific data rows.
[COMMAND_EXECUTION]: The skill enables the agent to interact with databases using structured commands.
The query_table tool allows the agent to generate and run SQL queries to manage and search product catalogs, price lists, and other tabular data.
[DATA_EXFILTRATION]: The skill facilitates network operations to external services.
The webhook and scrape_url tools allow the agent to communicate with arbitrary external endpoints, which could potentially be used to exfiltrate data if the agent is manipulated via prompt injection.

suvvy-mcp