The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The README.md file recommends an installation method (curl -sL skil.sh | sh) that downloads a script from a remote URL and executes it directly in the shell. This pattern allows for the execution of unverified code from an external domain.
[COMMAND_EXECUTION]: In SKILL.md, the workflow instructs the AI agent to execute shell commands (ls -la) to verify the success and structure of file system exports. While useful for verification, it grants the agent direct interaction with the host shell.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize dataset metadata and labels, which may contain untrusted data.
Ingestion points: Operations like dataset_summary, get_field_schema, and distinct('tags') in SKILL.md and HF-HUB-EXPORT.md read content from FiftyOne datasets that could be maliciously crafted.
Boundary markers: The instructions lack explicit boundary markers or directives for the agent to ignore potential instructions embedded within the dataset fields.
Capability inventory: The agent has the capability to write files to absolute paths on the local file system and upload data to the Hugging Face Hub using the execute_operator and FiftyOne SDK.
Sanitization: There is no evidence of sanitization or escaping of dataset content before it is processed by the agent to generate reports or Hugging Face dataset cards.

fiftyone-dataset-export