The Agent Skills Directory

[Secure Secret Management]: The skill explicitly instructs developers to use IConfiguration, environment variables, or user secrets instead of hard-coding API keys. It strongly recommends using DefaultAzureCredential for token-based authentication, which is a security best practice for managing cloud identities.
[Tool Approval and Side-Effect Handling]: The instructions include patterns for wrapping side-effecting functions (such as writes, purchases, or shell commands) in an ApprovalRequiredAIFunction. This ensures that dangerous operations require explicit human-in-the-loop confirmation rather than executing autonomously.
[Data Boundary Enforcement]: The documentation provides clear guidance on treating session state, chat history, and retrieved context as sensitive data. It advises storing this information in tenant-authorized storage with integrity protection and enforcing authorization boundaries outside the model.
[SSRF and Network Mitigations]: For declarative agents using HTTP handlers, the skill documents a mandatory allowlist mechanism (DefaultHttpRequestHandler) that denies all requests unless specific hosts are explicitly permitted by the application code.
[Sandboxing and Governance]: The skill identifies potentially sensitive integrations like LocalCodeAct and Tools.Shell, providing warnings that these should be sandboxed. It also mentions integration with Microsoft.Agents.AI.Purview for governance and DLP policy enforcement.

agent-framework-builder-dotnet