Static analysis reported a critical heuristic because this skill combines local Python execution, outbound network calls, and API credential access. Manual review found no evidence of malicious exfiltration or prompt injection; the confirmed risk is expected third-party API use with bearer credentials and user-supplied media prompts. Publish with clear warnings about external API transmission and local credential handling.
Authenticated requests send user inputs to external APIs
The client builds an Authorization bearer header from the configured API key and sends request bodies to MuleRouter or MuleRun endpoints. This is core functionality, but users must understand that prompts, media URLs, and task data leave the local environment.
The bearer token header and HTTP request call are directly visible. The behavior is legitimate for this API skill, but it is still a real data-sharing boundary.
The configuration loader imports python-dotenv, loads a provided file or current-directory .env file, and reads MULEROUTER_SITE and MULEROUTER_API_KEY. This is expected credential configuration, but it can expose credentials to the skill process.
The environment loading and API key lookup are explicit. I did not find evidence that unrelated environment variables are sent over the network.
Dynamic local module loading for model registration
The model listing workflow executes local provider package files through importlib so endpoints can register themselves. The provider list is fixed in the repository, so this is not arbitrary user-controlled import, but it is still local code execution.
The importlib exec_module call is present in the model listing script. The provider names are hardcoded rather than supplied by user input.
Markdown command examples are not runtime command execution
Static analysis flagged many fenced code examples and inline command references as shell execution. These are documentation examples in Markdown, not code paths that execute commands automatically.
The flagged locations are Markdown documentation blocks and inline examples. No runtime shell invocation is present at those locations.
Weak cryptography alerts are string false positives
The weak cryptographic algorithm cluster appears to match ordinary words such as description, model names, version strings, and seed parameters. I did not find cryptographic hashing, encryption, or signature code in the inspected implementation.
The cited lines contain descriptions and model parameters, not crypto APIs. No evidence of weak cryptographic usage was found in reviewed core files.
System reconnaissance alerts map to task metadata parsing
The low system reconnaissance alerts correspond to task response fields and model listing output, not host discovery or operating system inspection. No evidence found of system inventory collection.
The inspected lines parse API task IDs and optional site configuration. They do not enumerate files, users, processes, network interfaces, or host properties.