mercury-receipt-finder

Pass

Audited by Gen Agent Trust Hub on Jun 13, 2026

Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses highly sensitive data sources including the macOS Photos library (via Spotlight and AppleScript) and personal Gmail messages (via Google OAuth and API). This data is then processed and potentially uploaded to the Mercury platform. This behavior is consistent with the skill's primary stated purpose.
  • [COMMAND_EXECUTION]: The skill executes multiple local commands and scripts to perform its tasks:
  • Uses osascript to programmatically interact with the macOS Photos application and export media.
  • Uses mdfind to perform deep searches across the local filesystem via Spotlight.
  • Uses the mercury CLI for financial transaction management and file uploads.
  • Uses curl for direct network interaction with the Gmail API.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to the ingestion of untrusted external data.
  • Ingestion points: Processes raw email body text and subject lines from Gmail (Step 3b), and OCR/Vision text extracted from local images (Step 2e).
  • Boundary markers: No specific delimiters or instruction-isolation markers are used to separate untrusted content from the agent's internal logic.
  • Capability inventory: The agent has the ability to list/write financial transactions via the Mercury CLI, execute shell commands, and perform network operations.
  • Sanitization: There is no evidence of sanitization or filtering of the extracted text before it is processed by the model to make scoring or matching decisions.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 13, 2026, 12:52 PM