The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The reference.md file contains several instances of subprocess.run being used to execute external system commands for COLMAP and 3D Gaussian Splatting training. These implementations use dynamic paths (e.g., self.colmap_path) and execute local scripts (e.g., convert_to_web.py) which are not provided within the skill itself, creating a risk of arbitrary command execution if parameters are manipulated.
EXTERNAL_DOWNLOADS (MEDIUM): The allowed-tools section in SKILL.md permits the use of pip:* and python:* via Bash. This allows the agent to download and install arbitrary external packages from public repositories, which is a primary vector for supply chain attacks and remote code execution.
INDIRECT_PROMPT_INJECTION (LOW): The skill is configured to use WebFetch and mcp__firecrawl__firecrawl_search to ingest external data. This ingestion surface is vulnerable to indirect prompt injection where malicious instructions embedded in web pages or search results could attempt to override the agent's behavior.
DATA_EXPOSURE (LOW): The processing of video and image metadata, including GPS extraction as shown in reference.md, involves handling sensitive location data. While appropriate for the task, it represents a data surface that could be targeted for exfiltration if the agent's network tools are misused.

drone-inspection-specialist