The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit instructions to override safety and review protocols. In SKILL.md, it states: 'If the user's prompt includes "proceed without asking"... skip ALL review gates and proceed through the entire workflow without stopping.' This pattern encourages the agent to bypass confirmation steps designed for security and validation.
[COMMAND_EXECUTION]: The skill generates and executes MATLAB scripts (.m files) and registration files (sl_customization.m) using evaluate_matlab_code. While this is part of the intended functionality for authoring Model Advisor checks, the logic allows for dynamic code assembly based on user-provided natural language constraints without explicit sanitization steps.
[DATA_EXFILTRATION]: The skill performs network operations via evaluate_matlab_code to fetch license information (https://www.mathworks.com/content/dam/mathworks/license/pmrl/license.md). Although this targets a well-known service domain, the infrastructure for executing arbitrary MATLAB code provides a latent capability for data exfiltration if the prompt injection vectors are exploited to redirect output.
[DYNAMIC_EXECUTION]: The reference files (author-check.md, test-check.md) detail procedures for writing and immediately executing code and tests. The test-recipes.md specifically includes recipes for interactive API manipulation and auto-fix actions that modify model files and configuration parameters at runtime.

author-modeladvisor-checks