The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill configuration guide instructs users to run external packages using the uvx tool (e.g., uvx arxiv-mcp-server and uvx pubmedmcp@latest). This pattern involves fetching and executing code from remote registries at runtime without explicit verification of the source code.
[EXTERNAL_DOWNLOADS]: The skill references and encourages the download of third-party MCP servers from unofficial GitHub repositories (e.g., blazickjp/arxiv-mcp-server, grll/pubmedmcp, and 54yyyu/zotero-mcp) which are not managed by verified or trusted organizations.
[COMMAND_EXECUTION]: The skill uses Bash and curl to interact with local APIs and execute setup commands. Specifically, it provides templates for executing shell commands to query the Zotero API on the local host.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core function of reading and summarizing external research papers and web search results.
Ingestion points: The agent reads full-text papers via read_paper (ArXiv), PubMed queries, and general web content via WebFetch and WebSearch tools.
Boundary markers: No delimiters or defensive instructions are present to prevent the agent from following malicious instructions potentially embedded in the academic papers or web content being analyzed.
Capability inventory: The skill allows access to powerful tools like Bash, Write, Edit, and Task, which could be exploited if an external source successfully injects instructions into the agent's context.
Sanitization: There is no evidence of content filtering, escaping, or sanitization performed on the retrieved paper text before it is processed by the agent.

medical-imaging-review