reading-business-cards
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The skill performs network requests to the official Anthropic API (api.anthropic.com) to process image tiles. This is the primary mechanism for the skill's OCR functionality.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied images of business cards, which may contain adversarial text designed to influence the model's transcription behavior.\n
- Ingestion points: Image files are processed by scripts/prep_cards.py and scripts/extract_cards.py.\n
- Boundary markers: The transcription prompt in prompts/haiku_extract.md uses XML-style tags and strict rules to enforce JSON-only output and prevent data invention.\n
- Capability inventory: The skill scripts perform file writes (image tiles, manifest, and CSV output) and network API calls. These scripts are intended to be executed by the agent.\n
- Sanitization: The transcription script parses model output into a structured JSON format before writing to a CSV, though it does not explicitly sanitize the text content for malicious instructions.\n- [COMMAND_EXECUTION]: The skill requires the agent to execute local Python scripts (scripts/prep_cards.py and scripts/extract_cards.py) to perform image processing and data extraction tasks. These scripts use standard image processing and networking libraries.
Audit Metadata