The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a chat agent interface that ingests untrusted user data as part of its primary function.
Ingestion points: User input is received via context.activity.text in message handlers such as on_message, on_hello, and on_poem_message (interpolated into LLM prompts).
Boundary markers: The provided code snippets do not explicitly show the use of delimiters or 'ignore' instructions when processing user-supplied text.
Capability inventory: The agent possesses capabilities to call the Microsoft Graph API, access local MemoryStorage, and return streaming responses to end-users.
Sanitization: No explicit sanitization, filtering, or validation logic is present for incoming user messages before they are processed by the agent or forwarded to external APIs.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of several external Python packages to function.
Evidence: Instructions include pip install for microsoft-agents-hosting-core, microsoft-agents-hosting-aiohttp, microsoft-agents-activity, microsoft-agents-authentication-msal, and microsoft-agents-copilotstudio-client. These are official SDK packages from Microsoft, which is a recognized trusted organization. Other standard dependencies include python-dotenv, aiohttp, openai, and msal.
[SAFE]: The skill follows established security patterns for SDK integration.
Evidence: The implementation uses load_configuration_from_env and .env files for managing credentials like CLIENTSECRET and AZURE_OPENAI_API_KEY, which is a recommended practice to avoid hardcoding secrets. It also enforces authorization via jwt_authorization_middleware.

m365-agents-py