The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data retrieved via Qonto API calls, creating a surface for indirect prompt injection attacks where malicious content in transaction labels or invoices could attempt to influence the agent.
Ingestion points: The skill ingests data from list_transactions (labels), list_supplier_invoices (vendor names and text), get_organization (identity fields), and list_memberships (user data).
Boundary markers: The skill includes a dedicated 'Guardrails' section that explicitly instructs the agent to treat tool responses as untrusted data, never as instructions, and to ignore any embedded commands or system message claims within these fields.
Capability inventory: The skill possesses the Write tool for local file creation and multiple read-only MCP tools for accessing financial data.
Sanitization: The skill mandates strict HTML entity escaping (e.g., escaping &, <, >, ", and ') and prohibits the inclusion of scripts, event handlers, or external resources in its HTML output.
[DYNAMIC_EXECUTION]: The skill dynamically generates a self-contained HTML document at runtime based on external data.
The instructions require the output to be static and self-contained with inline styles only, mitigating risks associated with external resource loading or script execution.
It explicitly forbids executable code, automatic navigation, and embedded frames in the generated output.

qonto-rgpd-register