brainstorming

Warn

Audited by Gen Agent Trust Hub on May 2, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes shell scripts (start-server.sh, stop-server.sh) to manage a local Node.js server (server.cjs). This server is used to facilitate the 'Visual Companion' feature by serving HTML files and monitoring system processes.\n- [DATA_EXFILTRATION]: The visual companion server exposes files from the brainstorming session directory through a dedicated web endpoint (/files/). While intended for local use, the documentation in visual-companion.md suggests binding the server to 0.0.0.0 for remote access, which would make session mockups and interaction data accessible to the network.\n- [PROMPT_INJECTION]: The skill implements an indirect prompt injection surface. The agent is instructed to read browser interaction data from the state_dir/events file and incorporate it into its context. This unvalidated data, which includes user-controlled click events and selection text from the browser companion, enters the agent's context without sanitization or boundary markers. The agent possesses significant capabilities, such as writing design specifications to the filesystem and invoking downstream implementation skills, which could be influenced by injected instructions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 2, 2026, 12:08 AM