The Agent Skills Directory

[COMMAND_EXECUTION]: The README.md file instructs users to install the skill using curl -sL skil.sh | sh, which is an unverified remote script execution pattern that bypasses package manager safety checks.
[EXTERNAL_DOWNLOADS]: The skill uses the download_plugin tool in SKILL.md and QUALITY-CHECKS.md to fetch code from jacobmarks/image-quality-issues, an external repository not included in the trusted vendors list.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8). \n * Ingestion points: Untrusted data enters the agent context through dataset_summary, get_field_schema, and count_values which read dataset metadata, class labels, and annotation fields. \n * Boundary markers: No delimiters or instructions are provided to the model to ignore embedded instructions within the dataset content. \n * Capability inventory: The skill can execute arbitrary operators via execute_operator, modify dataset state with tag_samples, and execute background shell commands using fiftyone delegated launch. \n * Sanitization: There is no evidence of validation or filtering for data retrieved from the dataset before it is used in prompt interpolation.
[COMMAND_EXECUTION]: The skill executes shell commands (fiftyone delegated launch &) to manage background services, which involves subprocess spawning that could be exploited if service names were influenced by untrusted data.

fiftyone-dataset-curation