matlab-use-machine-learning-apps
Warn
Audited by Gen Agent Trust Hub on Aug 2, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill employs the 'evaluate_matlab_code' tool to run scripts and interact with MATLAB learner applications. It dynamically modifies the MATLAB search path using the 'addpath' command to include a directory of obfuscated P-code files.
- [COMMAND_EXECUTION]: The primary logic of the skill is contained within 22 '.p' files, such as 'AppController.p' and 'SessionController.p'. These are MATLAB P-code files that are pre-compiled and obfuscated, making it impossible to verify the exact MATLAB commands and operations being performed.
- [COMMAND_EXECUTION]: The skill provides functionality to generate executable MATLAB training scripts via the 'app.generateCode' method. This allows for the creation and execution of code strings that are dynamically constructed based on the current application state.
- [PROMPT_INJECTION]: The skill exposes an indirect prompt injection surface by ingesting untrusted user data, including tables and variable names, from the MATLAB workspace. 1. Ingestion points: 'Tbl', 'Y', and 'Response' arguments used in 'openApp' and 'importTestData'. 2. Boundary markers: There are no delimiters or 'ignore-instructions' warnings used to wrap the ingested user data. 3. Capability inventory: The skill can write to the file system (exporting models/results), generate new code, and execute arbitrary MATLAB via platform tools. 4. Sanitization: No sanitization or escaping of variable names or data content is performed before they are used in internal logic or code generation.
Audit Metadata