The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to perform shell command execution for environment setup and data retrieval. In references/01-meta-initialization.md, the agent is directed to create directories using mkdir -p. More significantly, references/00-research-standards.md instructs the agent to use curl -L -o file.pdf '<url>' to download documents discovered during web searches. Executing system commands with parameters derived from external, untrusted web content is a security risk, even with the provided advice to use single quotes and check for shell metacharacters.\n- [EXTERNAL_DOWNLOADS]: The research process involves downloading external files, particularly PDFs, from arbitrary URLs discovered at runtime (e.g., from corporate investor relations pages or guideline sites). The instructions for using curl in references/00-research-standards.md facilitate these downloads from non-predefined domains.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its ingestion of untrusted external content.\n
Ingestion points: Data enters the context via WebSearch, WebFetch, and MCP tools for PubMed and ClinicalTrials.gov as specified in SKILL.md and references/00-research-standards.md.\n
Boundary markers: No explicit delimiters or boundary instructions are provided to separate untrusted content, although references/00-research-standards.md directs the agent to distill findings into structured formats.\n
Capability inventory: The skill uses shell execution (curl, mkdir), file writing to waypoints, and file reading (Read).\n
Sanitization: Basic sanitization is suggested in references/00-research-standards.md, where the agent is told to avoid shell metacharacters and use single-quoting for dynamic URLs.

indication-dossier