The Agent Skills Directory

[DATA_EXFILTRATION]: The skill contains multiple hardcoded absolute file paths pointing to a specific user's OneDrive directory (e.g., C:/Users/AGasser/OneDrive/5 ANE CLAUDE work folder/mel_wiki/wiki/). While not a credential, this exposes the directory structure and username of the author's local environment.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes data from external web sources, which could contain malicious instructions designed to influence the agent's synthesis.
Ingestion points: Uses WebSearch, WebFetch, and specialized MCP servers (PubMed, Consensus) to pull information from the internet into the agent's context (e.g., targeting pubmed.ncbi.nlm.nih.gov, who.int, unfpa.org).
Boundary markers: Output is delimited with === EVIDENCE BRIEF === blocks, but there are no explicit instructions to ignore or sanitize instructions found within the fetched external research materials.
Capability inventory: The agent can spawn subagents, execute shell commands via Python, and read/write local files.
Sanitization: No explicit sanitization or filtering of external content is specified before the synthesis process.
[COMMAND_EXECUTION]: The skill invokes Python modules from a vendor-specific package (ane_package.bibliography.brief_push) to manage research citations.
[DYNAMIC_EXECUTION]: The workflow involves dynamically spawning a "verifier subagent" (Sonnet) with specific instructions to browse the web and validate citations drafted in the synthesis phase.
[EXTERNAL_DOWNLOADS]: The skill fetches data from well-known medical and institutional repositories including PubMed, WHO, UNFPA, and OpenAlex. These are recognized as legitimate professional services for research purposes.

researcher