The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The README.md file contains an installation instruction using curl -sL skil.sh | sh, which downloads and executes a remote script directly in the system shell without verification.
[DYNAMIC_EXECUTION]: The SKILL.md workflow includes Path C, which allows for the execution of models loaded from arbitrary remote GitHub repository URLs via the execute_operator command, creating a significant vector for running unverified code.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of missing Python dependencies using pip install triggered by the agent's interaction with the FiftyOne environment.
[COMMAND_EXECUTION]: The skill involves direct execution of shell commands for system configuration and software package management.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through data processed from datasets or instructions embedded in remote model sources.
Ingestion points: Sample data from FiftyOne datasets and content from user-provided remote GitHub repository URLs.
Boundary markers: No explicit delimitations or safety warnings are implemented to isolate untrusted data or model code from the agent's core instruction set.
Capability inventory: Includes capabilities for code execution via execute_operator, software installation via pip install, and local application serving via launch_app.
Sanitization: There are no verification steps described for the integrity of external repositories or the content of the datasets being processed.

fiftyone-dataset-inference