The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download and install a 'MATLAB MCP Core Server' and a toolkit from an external, unverified repository (github.com/matlab). This process involves executing remote setup scripts and binary installers (.mltbx files) which are not from a verified or trusted organization.
[COMMAND_EXECUTION]: The toolkit provides the evaluate_matlab_code tool, which allows the agent to execute arbitrary commands. This capability can be used to run OS-level commands through MATLAB's shell execution features (e.g., !command or system()), bypassing standard agent restrictions if manipulated.
[EXTERNAL_DOWNLOADS]: The installation instructions rely on multiple external downloads, including:
Cloning the toolkit repository from github.com/matlab/matlab-agentic-toolkit.git.
Fetching the MATLAB MCP Core Server from github.com/matlab/matlab-mcp-core-server.
Downloading binary installers from github.com/matlab/simulink-agentic-toolkit/releases.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through several vectors:
Ingestion points: Reads external files (run_matlab_file), imports data from files like CSV or Excel, and analyzes command output.
Boundary markers: None identified. There are no instructions for the agent to use delimiters or ignore instructions embedded in the processed data.
Capability inventory: Provides full MATLAB execution (evaluate_matlab_code), file system interaction, and potential network access via toolboxes.
Sanitization: Absent. The skill processes and executes code directly from external sources without validation or filtering.
[DYNAMIC_EXECUTION]: The core functionality relies on generating and executing MATLAB code at runtime via MCP tools, which is a form of dynamic execution that can be exploited if the generated code incorporates untrusted input.

matlab-agentic-toolkit