The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to load and run Python code from the literature-review/kernel.py file using exec(open(...).read()). This pattern involves dynamic code loading from local file paths that rely on environment-defined variables (e.g., <skills>), which is a mechanism used here for modular integration of related vendor skills.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external research data, creating a potential surface for indirect prompt injection.
Ingestion points: Technical descriptions and metadata from the OpenAlex API via search_openalex, and specific method sections extracted from PDFs using pdf-explore (pdf_outline, pdf_pages).
Boundary markers: The instructions lack specific delimiters or isolation warnings to the agent to prevent the execution of potentially malicious instructions embedded within retrieved research papers or metadata.
Capability inventory: The skill can execute dynamically loaded logic and interact with file exploration and experimental tools, which could be leveraged if a prompt injection occurs during the evaluation of external data.
Sanitization: There is no evidence of input validation, filtering, or sanitization of the content retrieved from external research sources before it is analyzed by the agent to make a 'Go/No-go' research decision.

research-ideation