The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run() in multiple orchestrator scripts to interact with the Azure CLI (az) and to chain execution of other Python scripts.
Evidence in scripts/auto_finetune.py, scripts/deploy_model.py, and scripts/chunk_and_generate.py shows the construction of command lines to fetch AAD tokens, deploy models via ARM, and launch data generation jobs.
In scripts/chunk_and_generate.py, arguments like --teacher and --recipe are passed directly into subprocess calls, which could be exploited if configuration artifacts are modified by an attacker.
[DYNAMIC_EXECUTION]: The script scripts/calibrate_grader.py utilizes exec() and compile() to load and execute Python code from a file provided by the user or the agent.
While the script includes a path validation check to ensure the grader file is within the current working directory, this mechanism allows for arbitrary code execution within the context of the agent's environment.
This is a core feature for Reinforcement Fine-Tuning (RFT) where custom grading logic is required, but it presents a risk if the agent is directed to execute a malicious grader file.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data (CSV, JSON, JSONL) for fine-tuning and evaluation, using LLM judges to score the outputs.
Ingestion points: Data enters the context via scripts/auto_finetune.py, scripts/evaluate_model.py, and scripts/quality_filter.py when loading datasets.
Boundary markers: Missing; prompt templates for the LLM judges (e.g., in scripts/evaluate_model.py) use simple string interpolation of model outputs without explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The skill has significant capabilities including file writing, network operations to Azure endpoints, and shell command execution.
Sanitization: There is minimal sanitization of the content from the processed datasets before it is sent to the evaluation models.
[EXTERNAL_DOWNLOADS]: Several scripts and documentation files reference external resources and APIs, all of which are hosted on official Microsoft and Azure domains.
scripts/deploy_model.py interacts with management.azure.com for ARM deployments.
scripts/content_safety_check.py sends text to the Azure Content Safety API.

azure-ai-fine-tuning