The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions are focused entirely on tax logic and screening criteria. No attempts to override safety filters, bypass guidelines, or extract system prompts were detected.
[DATA_EXFILTRATION]: No hardcoded credentials, sensitive file access, or unauthorized network operations are present. The skill provides reference links to official government domains (irs.gov) and recommends well-known local services (Bank of Israel) for currency conversion.
[OBFUSCATION]: Analysis of the markdown text and scripts found no evidence of Base64 encoding, zero-width characters, homoglyph attacks, or hidden payloads.
[COMMAND_EXECUTION]: The bundled Python script (scripts/screen_product.py) is a logic-only utility using standard library modules. It does not perform system calls, write files, or initiate network connections.
[INDIRECT_PROMPT_INJECTION]: The skill includes a defined attack surface by ingesting user descriptions of financial products. However, it lacks exploitable capabilities (such as file-writing or network exfiltration) and incorporates repeated legal disclaimers to ensure the agent remains within its defined role.
Ingestion points: User-provided product names, values, and purpose descriptions in Stage 1.
Boundary markers: Explicit instructions to state that the output is not a final classification at the start and end of the session.
Capability inventory: No file-write, no exec/eval, and no network-send capabilities are present in the skill's scripts.
Sanitization: The agent is instructed to apply statutory tests literally to the user's description, focusing on numerical and categorical data points.

us-person-israeli-investment-check