The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data in the form of student comments, which may contain instructions designed to manipulate the agent's behavior during analysis or reporting phases.
Ingestion points: The get_peer_review_comments tool in SKILL.md retrieves arbitrary text content authored by students.
Boundary markers: The skill instructions do not specify the use of delimiters or provide explicit "ignore embedded instructions" directives when processing student comment text.
Capability inventory: The agent has access to messaging tools (send_peer_review_inbox_messages, send_peer_review_followup_campaign) and file system write capabilities (extract_peer_review_dataset, generate_peer_review_report).
Sanitization: There is no evidence of input validation, filtering, or instruction-stripping for the ingested student feedback.
[DATA_EXFILTRATION]: The skill provides functional paths to export sensitive student data, including identifies and comments, from the Canvas API to the local file system.
Evidence: The extract_peer_review_dataset tool features a save_locally=true parameter, and generate_peer_review_report allows saving reports to a local file via save_to_file=true. While intended for instructor records, these parameters enable the permanent storage of academic data outside the managed Canvas environment.
[COMMAND_EXECUTION]: The skill utilizes a suite of Canvas-specific MCP tools to perform administrative actions.
Evidence: The skill orchestrates multiple tool calls (e.g., list_courses, get_peer_review_completion_analytics, send_peer_review_inbox_messages) to manage course state. Although these are managed via an MCP server, they represent significant administrative control over the target LMS environment.

canvas-peer-review-manager