The Agent Skills Directory

[OBFUSCATION]: The skill contains 18 binary-encoded MATLAB scripts (e.g., ecokg_decrypt.p, eco_snapshot.p, CodeMetricsFetcherSIL.p). This proprietary MATLAB format obfuscates the source code, making it impossible to audit the logic for malicious behavior, such as unauthorized file access or data exfiltration, during the optimization process.
[DYNAMIC_EXECUTION]: The skill uses a Knowledge Graph mechanism via ecokg_query and ecokg_detail to retrieve optimization suggestions. The ecokg_detail script returns an opt_reference field containing a "Full procedure" and "Commands Reference" that the agent is then instructed to follow. This results in the dynamic loading and execution of instructions and code snippets at runtime that are not present in the skill's static files.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection through several vectors:
Ingestion Points: The agent is instructed to read untrusted external data, including generated C/C++ source code, model metrics profiling reports, and a user-supplied configuration file (optimization_preferences.yaml).
Boundary Markers: The prompts used to delegate file-reading tasks to sub-agents (e.g., in Phase 2.5: User-Directed Function Targeting & Deep Analysis) lack explicit boundary markers or instructions to ignore embedded commands within the ingested content.
Capability Inventory: The skill utilizes powerful tools such as evaluate_matlab_code and run_matlab_file, which can execute arbitrary code within the MATLAB environment, combined with file system access and git repository manipulation (eco_snapshot.p).
Sanitization: There is no evidence of sanitization or strict schema validation for the data extracted from reports or source files before it is processed by the agent's logic.

simulink-optimizing-embedded-code