The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes external research paper metadata and abstracts from APIs such as Crossref, OpenAlex, and DataCite. This ingestion of untrusted external data combined with powerful capabilities creates an indirect prompt injection surface.
Ingestion points: Metadata is retrieved in agents/01-scout.md and summarized in agents/02-scribe.md.
Boundary markers: The instructions do not use specific delimiters or explicit 'ignore embedded instructions' warnings for the fetched external content.
Capability inventory: The agent can execute provided Python scripts (e.g., scripts/export.py which spawns subprocesses for pandoc) and has write access to the workspace.
Sanitization: There is no evidence of sanitization or validation of the content retrieved from external research databases before it is presented to the LLM.
[DYNAMIC_EXECUTION]: In agents/01-scout.md, the agent is instructed to execute a multi-line Python script provided via a shell heredoc (`python3
<<'PY'`) to deduplicate and merge JSON records. Although the script is statically defined within the prompt, this represents the execution of dynamically loaded code at runtime.

autonomous-research