The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because it is designed to process and summarize data from untrusted internal communication sources.
Ingestion points: The workflow involves retrieving real-time data from Slack channels, Gmail emails, Google Drive documents, and Calendar events as specified in SKILL.md and the reference templates in the references/ directory.
Boundary markers: The instructions do not explicitly require the use of delimiters or specific markers to isolate ingested content, which could lead the agent to follow instructions embedded within the retrieved data.
Capability inventory: The skill is limited to drafting, formatting, and presenting text to the user; it does not include capabilities for persistent storage, command execution, or network exfiltration.
Sanitization: No explicit sanitization, filtering, or validation steps for external content are provided in the skill instructions.

team-communications