paper-illustration

Warn

Audited by Gen Agent Trust Hub on Oct 24, 2026

Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the user via $ARGUMENTS in Step 1 and passes it through a multi-stage AI reasoning chain without boundary markers or sanitization.
  • Ingestion points: User input captured as $ARGUMENTS in SKILL.md.
  • Boundary markers: None identified; user input is directly used to generate instructions for subsequent Gemini stages.
  • Capability inventory: The skill has access to shell execution (bash), Python script execution (python3), and network operations (curl).
  • Sanitization: No evidence of input escaping, validation, or filtering of user-provided content before it influences the agent's logic flow.
  • [DYNAMIC_EXECUTION]: The skill uses inline Python blocks (python3 << PYTHON) to process JSON payloads and API responses. In Steps 2 and 4, it interpolates shell variables ($LAYOUT_REQUEST, $RENDER_PROMPT) containing user-derived data directly into triple-quoted Python strings ('''$VARIABLE'''). This allows an attacker to break out of the string literal using triple quotes and execute arbitrary Python code within the agent's execution environment.
  • [COMMAND_EXECUTION]: User-controlled input is interpolated into shell script templates across multiple steps (Steps 1, 2, 3, and 4). Without proper shell escaping, this creates a risk of command injection if the generated prompt content contains shell metacharacters like backticks or semicolons.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 24, 2026, 09:43 AM