bux-claude-agent

Fail

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONPRIVILEGE_ESCALATIONPERSISTENCEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill provides instructions to download and execute a shell script directly from a remote GitHub repository (browser-use/bux) using curl | bash with root privileges.
  • Evidence: curl -fsSL https://raw.githubusercontent.com/browser-use/bux/main/install.sh | sudo BROWSER_USE_API_KEY=bu_xxx bash
  • [PRIVILEGE_ESCALATION]: The skill makes extensive use of sudo to perform administrative tasks, including installing system services, modifying system-wide configuration files (like /etc/fstab), and executing commands as different users.
  • Evidence: Use of sudo for install.sh, systemctl commands, and sudo -u bux for running agent commands.
  • [PERSISTENCE]: The installation script sets up three background systemd services (bux-browser-keeper, bux-ttyd, and bux-telegram) to ensure the agent and its browser components remain active across reboots. It also includes a template for a cron job in /etc/cron.d/bux-morning-brief for scheduled task execution.
  • [INDIRECT_PROMPT_INJECTION]: The telegram_bot.py script creates a vulnerability surface by ingesting untrusted data from an external network source (Telegram) and interpolating it into a shell command.
  • Ingestion points: The handle_message function in telegram_bot.py reads update.message.text directly.
  • Boundary markers: None. The user input is passed as a raw string to the -p flag of the claude CLI.
  • Capability inventory: The agent has capabilities to browse the web via a persistent Chromium session, execute shell commands, and interact with the local filesystem in /home/bux.
  • Sanitization: The Python code does not implement sanitization or validation of the task variable before execution.
  • [DYNAMIC_EXECUTION]: The skill uses Python's subprocess module to dynamically execute shell commands constructed from potentially untrusted strings received over the network.
  • [COMMAND_EXECUTION]: The skill documentation encourages the agent to use SSH to access remote servers and execute configuration scripts, which involves the transmission of sensitive API keys and tokens in the command-line environment.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 14, 2026, 10:36 AM