The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes examples of prompt injection strings (e.g., 'Ignore previous instructions') within an 'Adversarial Testing' section in 'copilot-studio-byom.md'. These strings are provided for educational purposes to help developers test the robustness of their own implementations and do not constitute a threat within the skill's instructions.
[CREDENTIALS_UNSAFE]: The documentation consistently warns against hardcoding secrets, particularly in client-side code. It provides patterns for using Azure Key Vault, environment variables, and Managed Identities. Any API keys or tokens shown in code examples are generic placeholders (e.g., '[API key]', 'pk_live_xxx').
[EXTERNAL_DOWNLOADS]: The skill references well-known, official developer tools and libraries, including the Azure CLI and the official 'openai' Node.js package. These are standard industry dependencies used for their intended purpose.
[DATA_EXFILTRATION]: No patterns of unauthorized data exfiltration were detected. Network operations described in the patterns (e.g., calling Azure OpenAI endpoints) are legitimate and align with the stated purpose of the skill.
[DATA_EXPOSURE]: The skill provides surface areas for potential indirect prompt injection as it teaches how to interpolate user-provided data into AI prompts.
Ingestion points: User input is interpolated into prompt templates in 'power-automate-ai.md' (ticketDescription), 'code-apps-ai.md' (body.messages), and 'copilot-studio-byom.md' (userMessage).
Boundary markers: The basic code examples do not always include explicit boundary markers (like XML tags) around user input, which is a common but risky practice.
Capability inventory: The systems described (Azure Functions, Power Automate flows) have the capability to perform network requests and database updates based on AI responses.
Sanitization: Critically, the skill includes a 'Security Considerations' section in 'power-automate-ai.md' that explicitly instructs developers to 'Validate/sanitize user input before including in prompts' and 'Validate AI output before writing to Dataverse'. This educational approach mitigates the risk by informing the developer of the necessary safeguards.

azure-openai