The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied financial data, specifically annual gross revenue and expenses. While this introduces an ingestion surface for untrusted data, the skill mitigates risk by instructing the agent to pass this data to the calculadoraMEI.py script, which uses a strict numeric conversion method (_limpar_valor) that neutralizes potential text-based injection payloads.
[DYNAMIC_EXECUTION]: The agent is instructed to execute a local Python script (calculadoraMEI.py) to perform tax calculations. The script is static, provided within the skill's own package, and does not perform any unsafe operations such as loading remote code, using eval(), or spawning subshells with user input.
[METADATA_POISONING]: The SKILL.md documentation contains hardcoded absolute file paths referencing the author's local Windows development environment (e.g., C:/Users/mario/OneDrive/...). This represents a minor metadata leakage regarding the author's system structure but does not pose a security threat to users or the execution environment.
[SAFE]: The skill includes explicit security guardrails that direct the AI agent to ignore prompt injection attempts, refuse to provide illegal tax evasion advice, and block any attempts to execute unauthorized shell commands like curl or rm -rf.

mei-irpf