The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill is designed to process external, untrusted data from Microsoft 365 sources (emails, Teams messages, and meeting transcripts). This activity presents a surface for indirect prompt injection, where adversarial instructions contained within those signals could attempt to influence the agent's behavior.
Ingestion points: Data is retrieved via tools such as m365_list_emails, m365_list_chat_messages, and m365_get_transcript in SKILL.md.
Boundary markers: The skill uses metadata tags like [Quote], [Paraphrased], and [Review Required] to differentiate source content and highlight sections for user verification.
Capability inventory: The agent's capabilities are focused on local file operations within the institutional-knowledge directory; no tools for arbitrary code execution or external network exfiltration are utilized.
Sanitization: The instructions include heuristics for excluding sensitive categories (such as HR or health data) and filtering out automated system messages to reduce the exposure to potentially problematic content.

institutional-knowledge