The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a pipeline that ingests research papers from external, untrusted sources and uses them to automatically generate executable skill packages. This creates a surface for indirect prompt injection if a source paper contains adversarial text designed to override agent behavior during the generation or execution of the resulting skill.\n- Ingestion points: SKILL.md (Phase 1) describes fetching papers via arXiv IDs, web URLs, or direct text input.\n- Boundary markers: The instructions lack explicit boundary markers or directives to the agent to disregard instructions embedded within the source paper's text during analysis.\n- Capability inventory: The generated output includes SKILL.md instructions and evals/cases.yaml (containing test assertions), which are meant to be executed by the agent environment.\n- Sanitization: No sanitization or filtering logic is present to validate the content extracted from papers before it is used to generate the new skill's methodology.\n- [EXTERNAL_DOWNLOADS]: Fetches research papers from arXiv and other web URLs. Documents from these sources are used as the primary source of truth for generating new code and instructions. While arXiv is a well-known service, the specific documents are user-contributed and thus untrusted.

paper-to-skill