The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external scientific sources, creating a surface for indirect prompt injection.\n
Ingestion points: External literature is retrieved via WebFetch (targeting PubMed) and WebSearch results as described in the Step 2 workflow of SKILL.md.\n
Boundary markers: The instructions do not define explicit delimiters or 'ignore' directives to prevent the agent from obeying instructions potentially hidden within the fetched scientific papers.\n
Capability inventory: The agent is authorized to use Bash for command execution and has Write/Edit permissions, which increases the impact if an injection were to occur.\n
Sanitization: There is no requirement or logic provided to sanitize or validate the content retrieved from external searches.\n- [COMMAND_EXECUTION]: The skill relies on the execution of shell commands for core reporting and visualization tasks.\n
Evidence: SKILL.md instructs the agent to run 'python scripts/generate_schematic.py' to create required figures.\n
Evidence: The workflow requires a multi-step LaTeX compilation process using 'xelatex' and 'bibtex'.\n- [EXTERNAL_DOWNLOADS]: The skill fetches data from external scientific services to ground its hypotheses.\n
Evidence: Step 2 of the workflow uses WebFetch to access the National Center for Biotechnology Information (NCBI) PubMed database.\n
Evidence: The workflow uses WebSearch to find recent papers and preprints from across the web.\n- [REMOTE_CODE_EXECUTION]: The skill references an external Python script ('scripts/generate_schematic.py') that is not included in the provided file list, making its internal logic and safety unverifiable.

hypothesis-generation