The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates how to build agents that can ingest external data through retrieval-augmented generation (RAG) and local file system access. This pattern introduces a vulnerability surface where malicious instructions hidden in external data could attempt to influence the agent's behavior.
Ingestion points: FileReadExecutor in references/shared-states.md and AzureAISearchCollection in references/agent-rag.md allow the agent to process external content.
Boundary markers: The skill encourages the use of human-in-the-loop approvals (approval_mode="always_require") to mitigate risks associated with sensitive tool calls.
Capability inventory: The skill utilizes powerful tools including HostedCodeInterpreterTool for code execution, FileReadExecutor for file system access, and HostedMCPTool for network operations.
Sanitization: While the skill mentions security middleware, the provided code snippets do not show explicit sanitization or delimitation of external inputs.
[EXTERNAL_DOWNLOADS]: The skill references official installation commands and source code repositories for the Microsoft Agent Framework and its dependencies. All external references target trusted sources.
Evidence: pip install agent-framework-core, pip install agent-framework-azure-ai, and git clone https://github.com/microsoft/agent-framework.git.
[COMMAND_EXECUTION]: The skill includes documentation for the HostedCodeInterpreterTool, which allows agents to generate and execute Python code in a sandboxed environment for analysis purposes.
[SAFE]: Code snippets correctly use placeholders like your-api-key-here, InstrumentationKey=..., and <myresource> for sensitive values. The skill also recommends using .env files and environment variables for secret management, which is a standard security best practice.

microsoft-agent-framework