The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The utility script scripts/new_pipeline.py performs filesystem operations using unsanitized user input.
Evidence: The function create_pipeline takes a name argument and uses Path(name).mkdir(parents=True) and path.write_text(content) to create a project structure.
Risk: Without sanitization, a user could provide a name containing path traversal sequences (e.g., ../../) to write files outside of the intended workspace.
[INDIRECT_PROMPT_INJECTION]: The skill defines a 'Pipeline Review' mode that analyzes external code, which creates a surface for indirect instructions to influence agent behavior.
Ingestion points: The skill accepts code snippets for audit and review via its primary instruction loop in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded prompts within the reviewed code are provided.
Capability inventory: The agent can generate executable code, design architectures, and recommend system-level changes based on its analysis.
Sanitization: The skill does not include steps to sanitize or escape the content of the pipelines it reviews.

data-pipelines