The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by ingesting and processing untrusted data from multiple sources without security boundaries.
Ingestion points: The workflow described in SKILL.md ingests user-provided inputs such as disease_name, drug_name, and moa_term. It also processes data returned from external tools and research databases including PubMed, ChEMBL, DrugBank, and FDA FAERS.
Boundary markers: The instructions do not define delimiters (like XML tags or markdown blocks) or explicit 'ignore embedded instructions' warnings for the data fetched from external research tools before it is processed by the agent.
Capability inventory: The skill includes access to various network-enabled research tools (e.g., PubMed_search_articles, ChEMBL_search_drugs, ClinicalTrials_search) which allow for complex data retrieval operations.
Sanitization: The skill lacks logic to sanitize or validate the content retrieved from external databases before the agent is instructed to analyze, rank, and summarize the findings.

tooluniverse-drug-repurposing