The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone a repository from an untrusted GitHub account (LanceZPF/agent-as-a-router.git) and download dataset assets from an untrusted Hugging Face user (Lance1573). These sources are not recognized as trusted organizations or well-known services.
[REMOTE_CODE_EXECUTION]: The instructions direct the user to execute multiple Python scripts from the cloned repository (scripts/run_id.py, scripts/run_pipeline.py, scripts/download_hf_assets.py) which allows for arbitrary code execution from an unverified source.
[COMMAND_EXECUTION]: The skill implements and documents a "verifier" pattern that uses subprocess.run to execute code generated by Large Language Models (LLMs). This creates a direct path for a model to execute malicious shell commands on the host system if the generated code contains harmful payloads.
[PROMPT_INJECTION]: The skill processes untrusted external data, including task prompts and model-generated responses, without implementing sanitization or boundary markers.
Ingestion points: Task prompts from files like my_tasks.jsonl and code solutions generated by external model APIs.
Boundary markers: The provided implementation examples do not include delimiters or instructions to ignore embedded commands within the processed data.
Capability inventory: The system has access to subprocess.run for shell execution and network access via the openai client library.
Sanitization: There is no evidence of escaping, validation, or filtering of model-generated code before it is passed to the pytest runner or shell.

agent-as-a-router-coding