The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests experimental data from external files (CSV, JSON, pickle) via the DATA_PATH provided in Phase 1. This data is processed and its values are interpolated into a drafting phase for a research report in Phase 4. This creates a surface for indirect prompt injection if the data source contains malicious instructions designed to hijack the agent's writing or quality check phases.
Ingestion points: File reading logic in phases/01-data-loading.md using user-provided paths.
Boundary markers: Absent. The skill logic does not implement specific delimiters or 'ignore' instructions for the data content before interpolation.
Capability inventory: The skill has access to Bash, Write, and Edit tools as defined in SKILL.md.
Sanitization: Phase 1 includes consistency and type checks, but lacks content-based sanitization or escaping for Natural Language instructions.
[DYNAMIC_EXECUTION]: The skill documentation and implementation support the loading of Python pickle (.pkl) files as an experimental data format. Deserializing pickle data from untrusted sources is a known critical security risk as it can lead to arbitrary code execution.
Evidence: phases/01-data-loading.md lists .pkl in the supported file extensions for Step 1.1 and explicitly mentions loading pickle objects.

scholar-experiment