The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill contains calls to .eval() in tips_zones.py. These are PyTorch method calls (e.g., model_image.eval()) used to set neural network modules to evaluation mode. These are distinct from the built-in Python eval() function and do not pose a risk of arbitrary code execution.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs the user to clone the google-deepmind/tips repository from GitHub. This is a trusted source used for model architecture definitions necessary for the skill's primary functionality.
[COMMAND_EXECUTION]: The SKILL.md file provides standard shell commands for installing machine learning dependencies via pip and cloning the model repository via git. These are standard environment setup procedures for research-oriented skills.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-supplied natural language labels as prompts for image segmentation.
Ingestion points: Text prompts are ingested through the targets, edges, and distractors arguments in tips_zones.py.
Boundary markers: No explicit instruction boundary markers or "ignore" warnings are used for these inputs.
Capability inventory: The skill performs file system operations (creating an __init__.py file) and dynamic path manipulation (sys.path.insert) to facilitate model loading in tips_zones.py.
Sanitization: No sanitization is performed on input labels before they are tokenized for the model.
Risk Assessment: The risk is assessed as safe because the prompts are only used to generate cosine similarity heatmaps for bounding box detection, which does not influence the agent's logic or execution environment.

detecting-tips-zones