The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by ingesting untrusted data from local files and URLs via vstash add. This data is then used as context for the vstash ask and vstash chat commands. The instructions do not specify the use of boundary markers to separate retrieved context from agent instructions.
Ingestion points: The vstash add command in SKILL.md and references/cli-reference.md accepts arbitrary file paths and web URLs.
Boundary markers: No specific boundary markers or instruction-ignoring prompts are defined for the retrieval phase.
Capability inventory: The skill provides tools for file reading, network-based data ingestion, semantic search, and LLM-based questioning.
Sanitization: No sanitization or filtering of the ingested document content is described before it is provided to the agent.
[COMMAND_EXECUTION]: The skill uses a set of CLI tools (vstash) to manage local data. An agent could be instructed to use vstash add on sensitive system files or configuration directories, effectively moving sensitive data into the searchable memory of the AI, where it could be exfiltrated or misused.
[EXTERNAL_DOWNLOADS]: The vstash add command supports URLs, which allows the skill to fetch and process content from remote servers. This facilitates the ingestion of external data that has not been verified for safety or malicious instructions.

cli-vstash