The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill performs exhaustive research by fetching and processing data from a wide variety of untrusted sources, including web search results, research papers, documentation sites, and community forums. While the skill implements best practices such as boundary markers for LLM output and sanitization of file paths, the ingestion and interpolation of large amounts of external content into the agent context constitutes a significant surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through WebSearch, WebFetch, and API-based research adapters in files like rlm_runner.py and workflow.py.
Boundary markers: The skill uses clear markers (STRUCTURED_OUTPUT_START and STRUCTURED_OUTPUT_END) to separate the findings from structured metadata, reducing the risk of content being misinterpreted as instructions.
Capability inventory: The skill has significant capabilities, including shell command execution for sub-process orchestration (via subprocess.Popen in rlm_orchestrator.py) and filesystem access for writing findings.
Sanitization: Input used to create directories or files is sanitized in rlm_orchestrator.py using the _slugify_name function, which employs a restrictive whitelist regex ([^a-zA-Z0-9._-]) and prevents directory traversal.
[EXTERNAL_DOWNLOADS]: The research workflow involves downloading and querying data from several well-known technical and academic platforms, such as Arxiv, Semantic Scholar, GitHub, and Wikipedia. These operations are documented as part of the skill's primary research objective and target established, reputable services.

deep-research