close-month

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • Potential for Indirect Prompt Injection: The skill is designed to ingest and process transaction data from external sources such as QuickBooks, PayPal, Stripe, and Square. In scenarios where external data contains unstructured text (like transaction descriptions or memos), there is a risk that this data could influence the agent's behavior. The skill mitigates this through explicit human-in-the-loop requirements, such as waiting for the owner to triage flagged items and requiring confirmation for deletions.
  • Data Ingestion Points: Financial data is pulled from QuickBooks and external payment processors (Step 1).
  • Boundary Markers: The workflow does not explicitly define delimiters for external text strings to isolate them from agent instructions.
  • Capability Inventory: The skill is configured to use tools like Bash and WebFetch, which are necessary for accessing APIs and saving reconciliation packets to local or cloud storage.
  • Sanitization: The instructions do not specify a sanitization process for external transaction data before it is cross-referenced or narrated.
  • File Management and Export: The skill automates the creation and saving of financial reports (Step 4). While this involves writing files to local or cloud directories, the activity is consistent with the primary purpose of accounting reconciliation and is limited to specific formats like .xlsx and .pdf.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 11:38 PM