The Agent Skills Directory

Dynamic Code Execution: The calibrate_grader.py script uses exec() and compile() to load and run Python logic from a local file provided via the --grader argument. This is used to run custom grading functions for Reinforcement Fine-Tuning (RFT). While this is a documented feature, it allows for the execution of arbitrary Python code from the specified file.
Indirect Prompt Injection Surface: The skill harvests production traces and agent insights (in trace.md and insights.md) which can include untrusted data such as user messages or tool outputs. This creates a potential surface for indirect prompt injection when these traces are reviewed or used for agent optimization.
Command Execution: Several scripts (e.g., deploy_model.py and environment verification scripts) use subprocess.run() to execute Azure CLI (az) and Azure Developer CLI (azd) commands. This is used for legitimate management tasks like resource provisioning and token retrieval but involves spawning shell processes.
External Resource References: The skill fetches templates and samples from official Microsoft GitHub repositories (e.g., github.com/Azure-Samples/ and github.com/microsoft-foundry/). These are recognized as trusted vendor sources used for scaffolding agent projects.

microsoft-foundry