The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from WebSearch results and the user's Zotero library to generate research proposals.
Ingestion points: Phase 2: Literature Collection uses WebSearch and mcp__zotero tools to gather research data from the internet and the user's document library.
Capability inventory: The skill has Write and Edit permissions to generate and save the proposal as a Markdown file in the user's directory.
Boundary markers: The instructions lack explicit delimiters or "ignore instructions" warnings for text retrieved via WebSearch or Zotero content retrieval tools.
Sanitization: There is no evidence of sanitization or filtering for the external content before it is processed by the LLM for content generation.
Impact: An attacker could poison web search results or literature metadata to influence the generated proposal's content, although the risk is limited to document generation behavior.
[SAFE]: The skill focuses on academic automation and uses standardized MCP tools for Zotero integration. It does not contain any patterns for credential exfiltration, persistence, or privilege escalation. All external data processing is aligned with the stated primary purpose of the skill.
[SAFE]: No obfuscation, prompt injection overrides, or dangerous command execution patterns were found. Conversion instructions (e.g., using pandoc) are provided as educational documentation for the user and are not executed autonomously by the agent.

research-proposal