The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the auto_label command-line tool within a Docker container to perform its primary function of image grounding. This is the intended and documented behavior.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill includes instructions for installing vLLM and pulling Docker images from nvcr.io and the vllm-project repository. These references are to well-known and reputable industry resources.
[DATA_EXPOSURE_AND_EXFILTRATION]: Instructions for providing API keys (e.g., GOOGLE_API_KEY) utilize environment variables or configuration files, which is a secure method for managing sensitive credentials. No hardcoded secrets were detected.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided captions and images for processing by a VLM. While this is a standard data-processing workflow, it represents a surface where external content could theoretically influence model behavior.
Ingestion points: data.input_jsonl in SKILL.md and references/configuration.md.
Boundary markers: No explicit delimiters or instructions are used to separate user data from the VLM's task instructions.
Capability inventory: The pipeline can execute the auto_label CLI, make network calls to VLM APIs, and write annotation files and checkpoints to the local filesystem.
Sanitization: No specific sanitization or filtering of the input caption field is described in the documentation.

tao-generate-image-grounding