The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's core instructions and example scripts demonstrate the use of the eval() function to implement calculation tools. Specifically, SKILL.md provides an example of a calculate tool that directly executes user-supplied strings, and examples/02-tool-integration.py uses eval() with limited protections. Executing arbitrary strings via eval() is a high-risk pattern that can lead to remote code execution.
[DYNAMIC_EXECUTION]: The C# example in examples/05-tool-integration.cs employs DataTable.Compute() to evaluate expressions. As noted in the code's own comments, this method can be leveraged to execute arbitrary expressions if input is not properly sanitized.
[COMMAND_EXECUTION]: The reference documentation in reference/04-tools-functions.md includes a run_shell_command tool example using subprocess.run() with shell=True. Providing agents with shell access, especially with the potential for command injection via unsanitized arguments, poses a significant security risk.
[INDIRECT_PROMPT_INJECTION]: The skill documents how to build agents that ingest data from external sources (such as RAG providers or database tools) and use that data to drive high-privilege operations (file writes, command execution). The examples provided do not demonstrate the use of boundary markers or input validation, which creates a vulnerability to indirect prompt injection where instructions embedded in external data could hijack agent control flow.
[EXTERNAL_DOWNLOADS]: The skill recommends installing the agent-framework-core Python package and Microsoft.Agents.AI .NET package. These appear to be legitimate packages from a well-known organization and are documented neutrally.

microsoft-agent-framework