The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill requires the agent to ingest and analyze external data from academic papers and code repositories (Steps 2A and 2B in references/decomposition-prompts.md). This process creates a potential attack surface where an adversary could embed malicious instructions within a paper's text or a repository's README to influence the agent's behavior.
Ingestion points: The workflow explicitly directs the agent to search for and extract content from academic papers and codebases (SKILL.md and references/decomposition-prompts.md).
Boundary markers: The prompts do not include explicit delimiters or instructions for the agent to ignore potentially malicious embedded content within the source materials.
Capability inventory: The skill itself focuses on extraction and documentation (producing a JSON knowledge base), but it is explicitly linked to downstream tools like experiment-code which may possess higher-privilege capabilities.
Sanitization: There is no requirement for the agent to sanitize or validate the integrity of the LaTeX formulas or code snippets extracted from external sources before including them in the final knowledge entry.

atomic-decomposition