The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to run Python code via python3 -c using a template where user-provided values for INPUT_PATH, OUTPUT_PATH, and especially TEXT_PROMPT are substituted directly into the code string. An attacker can provide a specially crafted prompt containing single quotes and Python code (e.g., ', __import__('os').system('...'), ') to break out of the string literal and execute arbitrary commands on the host system.
[COMMAND_EXECUTION]: The skill utilizes the find command with user-supplied arguments ($1) to resolve file paths. While the variable is quoted in the instructions, the pattern of using user input to construct shell commands without explicit sanitization steps creates a risk of argument injection or unexpected behavior if the agent improperly formats the command.
[EXTERNAL_DOWNLOADS]: The skill requires downloading pre-trained AI models from Hugging Face at runtime. While Hugging Face is a well-known service, the execution of these models involves loading serialized weights and potentially untrusted code dependencies (e.g., via geoai-py and torch), which constitutes a remote code execution surface.
[EXTERNAL_DOWNLOADS]: The skill explicitly directs the user to install third-party packages geoai-py, torch, and torchvision. While these are established libraries in the geospatial and AI domains, they increase the attack surface of the environment.

detect-objects