code-game-trailer

Warn

Audited by Gen Agent Trust Hub on Nov 14, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The scripts/director-runtime.js file contains a utility function ev = (expr) => (typeof expr === 'string' ? (0, eval)(expr) : expr) which uses eval() to execute strings provided in the director.json configuration file. This is used to resolve references to the game's renderer, camera, and internal state at runtime.
  • [COMMAND_EXECUTION]: Multiple scripts (capture.ts, cut.ts, check.ts, beats.ts) use Bun.spawn to execute shell commands for ffmpeg, ffprobe, magick, and chrome. These operations are used for frame capture, audio analysis, and video encoding.
  • [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the director.json file, which serves as a control surface for the agent's actions.
  • Ingestion points: The director.json file, specifically fields such as ready, probe, setup, hooks, and camera.space, which are treated as executable JavaScript expressions.
  • Boundary markers: None identified; expressions are evaluated directly within the browser session.
  • Capability inventory: The skill possesses capabilities for file system writes via Bun.write, arbitrary command execution via Bun.spawn, and browser interaction through a custom CDP client.
  • Sanitization: There is no sanitization or validation of the expressions provided in the configuration file before they are passed to eval().
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 14, 2026, 09:49 PM