order-form-cro

Warn

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill's preamble and completion protocols execute multiple local shell commands including grep, sed, mkdir, touch, and ls. It also invokes custom scripts like rmbc-update-check, rmbc-analytics, and rmbc-workspace located in the skill's root or user's home directory.
  • [DATA_EXFILTRATION]: Executes a telemetry script (rmbc-analytics) that transmits data about skill usage and active product names to a remote server.
  • [EXTERNAL_DOWNLOADS]: Includes instructions to open external URLs (YouTube) and references a process for checking and applying updates from external sources via rmbc-update-check.
  • [PROMPT_INJECTION]: The skill ingests untrusted user data (product_name, target_audience, etc.) into its generation logic without explicit boundary markers or sanitization, creating a surface for indirect prompt injection. Additionally, it contains meta-instructions to parse and follow logic from external files like CHANGELOG.md or upgrade scripts found in rmbc-upgrade/SKILL.md.
  • Ingestion points: Input parameters such as product_name, price_point, and target_audience defined in SKILL.md.
  • Boundary markers: None present; data is interpolated directly into instructions.
  • Capability inventory: Shell script execution (bin/rmbc-*), file system read/write access (~/.rmbc-skills/), and external URL launching.
  • Sanitization: No escaping or validation of external content is specified before processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 11, 2026, 01:37 PM
Security Audit — agent-trust-hub — order-form-cro