matlab-engineer-tabular-features

Warn

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: MEDIUMOBFUSCATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [OBFUSCATION]: The skill includes eight MATLAB P-code files (.p) in the scripts/ directory. P-code is an obfuscated, pre-parsed format for MATLAB scripts that is not human-readable. This obscures the implementation logic for critical components such as feature ranking and mutual information analysis.
  • Evidence: Files include expandBlockMap.p, lambdaSequence.p, lassoFeatureRanking.p, mrmrFeatureRanking.p, mutualInformationPanel.p, ncaFeatureRanking.p, oobFeatureRanking.p, and scoreFeaturePanel.p.
  • [DYNAMIC_EXECUTION]: The skill performs dynamic code generation by writing new MATLAB function files (.m) to the user's local disk during the delivery phase.
  • Evidence: The script scripts/writeInferenceScript.m uses fprintf to construct and emit the source code for an inference function (fe_transform_.m). This function is designed to load and execute state from a companion .mat file, which may include function handles if the domain-routed path was used.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process arbitrary tabular data provided by the user, which represents a potential attack surface for indirect prompt injection.
  • Ingestion points: Data is ingested into the MATLAB environment through the RawTbl input during the Intake phase, as described in references/intake.md.
  • Boundary markers: There are no explicit markers or validation steps to ensure that data content does not contain instructions targeting the LLM.
  • Capability inventory: The skill has the capability to write files to disk (fprintf, save, exportgraphics) and execute function handles (validateRecipe.m), providing a path for malicious data to influence system state.
  • Sanitization: The skill uses matlab.lang.makeValidName to sanitize dataset names used in file paths, but it does not perform sanitization on the data values themselves.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 19, 2026, 01:51 PM
Security Audit — agent-trust-hub — matlab-engineer-tabular-features