The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection because it is designed to ingest and process data from untrusted external sources.
Ingestion points: The workflow in SKILL.md explicitly directs the agent to fetch content from user-provided links, files, and integrations such as Slack, Teams, Google Drive, and SharePoint.
Boundary markers: The instructions do not define clear boundaries or provide directives to the agent to ignore instructions that might be embedded within the documents or messages being read.
Capability inventory: The skill utilizes tools like create_file and str_replace to generate and modify files in the user's workspace, and it has the capability to search connected tools.
Sanitization: No sanitization or validation mechanisms are described to filter out potential malicious commands hidden within the external context before it is used to drive the drafting and refinement process.

doc-coauthoring