drone-inspection-specialist

Warn

Audited by Gen Agent Trust Hub on Mar 5, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes the colmap binary and external Python scripts such as train.py, render.py, and convert_to_web.py using subprocess.run to perform 3D reconstruction tasks. These commands are constructed dynamically within the GaussianSplattingReconstructor and GaussianSplatInspectionTool classes.\n- [DATA_EXFILTRATION]: The FireAlertSystem class includes a send_to_dispatch method that utilizes requests.post to transmit alert payloads—including geolocation coordinates, timestamps, and temperature data—to a remote endpoint defined at runtime.\n- [PROMPT_INJECTION]: The skill uses mcp__firecrawl__firecrawl_search and WebFetch to ingest data from external websites, creating a surface for indirect prompt injection. Ingestion points: firecrawl_search, WebFetch in SKILL.md. Boundary markers: None identified in the processing logic. Capability inventory: Bash, Write, Edit, and subprocess calls. Sanitization: No sanitization or filtering of retrieved web content is implemented before it enters the agent's context.\n- [EXTERNAL_DOWNLOADS]: The reconstruction pipeline generates an HTML viewer that fetches JavaScript libraries from cdn.jsdelivr.net (Three.js). Additionally, the skill's reliance on the ultralytics library for YOLO models often leads to the automatic download of pre-trained model weights from official repositories.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 5, 2026, 01:20 PM