The Agent Skills Directory

[SAFE]: The skill is configured with zero network permissions and strictly read-only tool access, ensuring that financial data cannot be exfiltrated or modified.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transaction data (counterparty, label, and amount) ingested via list_transactions and list_supplier_invoices in SKILL.md. While explicit boundary markers are absent for this ingested data, the capability inventory is limited to the Read tool with no network, shell, or write access enabled. The skill instructs sanitization by masking IBANs and personal names in reports, mitigating potential data exposure risks. Given the lack of dangerous capabilities, the injection surface is assessed as non-exploitable.

qonto-tax-radar