The Agent Skills Directory

[SAFE]: No malicious patterns detected. The instructions follow standard operational procedures for research management and project orchestration. There are no indications of prompt injection, credential exfiltration, or unauthorized command execution.
[NO_CODE]: The skill consists entirely of markdown instructions and organizational frameworks. It does not include any scripts (Python, JavaScript, etc.) or executable components that could pose a direct security risk to the host environment.
[INDIRECT_PROMPT_INJECTION]: While the skill defines a workflow for synthesizing literature and benchmarking data from external sources, it functions as a high-level manager and does not directly implement data ingestion or processing logic that could be exploited. The design relies on the inherent safety guardrails of the sub-skills it orchestrates.
Ingestion points: External research literature and competitive leaderboard data (Phase 3 and 4).
Boundary markers: The skill suggests using a 'Research Mission Brief' and 'Thread Registry' to maintain boundaries, though these are for organizational purposes rather than security sandboxing.
Capability inventory: The skill manages sub-skills but does not perform file writes, network requests, or system commands itself.
Sanitization: Sanitization is handled by the underlying LLM and the sub-skills performing the actual data retrieval.

principal-scientist