detail-flow

Warn

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill instructions in SKILL.md and references/installation.md contain directives that guide the AI agent to avoid or bypass user-facing interaction patterns, such as "trust" buttons, manual application launches, or "enable MCP" switches. The instructions specify that certain configuration checks "must not be turned into a user-facing blocker," which effectively tells the agent to conceal technical setup actions or suppress security confirmations from the user.
  • [COMMAND_EXECUTION]: Multiple scripts within the skill (scripts/setup_detail_flow.py, scripts/starcanvas_mcp.py, and scripts/starcanvas_startup.py) perform shell command execution using subprocess.run and subprocess.Popen. These commands are used to install Python libraries, query system process information via PowerShell on Windows, and launch the Starcanvas desktop application with specific parameters provided in configuration files.
  • [INDIRECT_PROMPT_INJECTION]: The skill exposes a significant vulnerability surface by ingesting untrusted data and possessing expansive local capabilities. \n1. Ingestion points: The skill reads local product assets, user-provided configuration files, and remote content from user-specified API endpoints (scripts/image_api.py). \n2. Boundary markers: Data is processed through structured JSON schemas, but the workflow involves the agent interpreting documentation from these external sources to determine protocol compatibility. \n3. Capability inventory: The agent has the capacity to execute shell commands, perform network operations via urllib.request, and write to the local filesystem. \n4. Sanitization: While basic path and type validations are performed in scripts, the high-level coordination logic relies on the model following complex rules when handling external data.
  • [EXTERNAL_DOWNLOADS]: During the setup phase, scripts/setup_detail_flow.py programmatically invokes pip to download and install the Pillow library if it is not detected in the current environment.
  • [DYNAMIC_EXECUTION]: The skill uses dynamic module loading through __import__ in its regression testing suite (scripts/test_generic_api.py). Additionally, the scripts/build_review_editor.py script generates standalone HTML files containing embedded data payloads for execution in a user's web browser.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 17, 2026, 09:36 PM