browser-harness

Warn

Audited by Gen Agent Trust Hub on Oct 29, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPERSISTENCEDATA_EXFILTRATION
Full Analysis
  • [DYNAMIC_EXECUTION]: The browser-harness command line tool executes Python code passed as a string via the -c flag. This allows for arbitrary code execution on the local machine where the script is invoked.\n- [REMOTE_CODE_EXECUTION]: The installation guide instructs the agent to clone an external repository from GitHub (browser-use/browser-harness) and install it using uv tool install. This pattern involves downloading and installing executable code from a third-party source.\n- [INDIRECT_PROMPT_INJECTION]: The skill has a high surface area for indirect prompt injection as it retrieves and processes arbitrary content from web pages, including authenticated sessions.\n
  • Ingestion points: Untrusted data enters the agent context through page_info() calls, DOM extraction via js() snippets, and visual screenshots of web pages.\n
  • Boundary markers: There are no explicit instructions or delimiters used to warn the agent to ignore instructions embedded within the extracted web content.\n
  • Capability inventory: The skill possesses extensive capabilities including shell execution, local file writing, and network operations via CDP.\n
  • Sanitization: No sanitization or filtering is applied to the extracted DOM text or page data before the agent processes it.\n- [COMMAND_EXECUTION]: The skill documentation provides multiple shell commands for installation and maintenance, including git clone, uv tool install, mkdir, ln -sf, and osascript to automate the browser UI.\n- [PERSISTENCE]: The installation process involves modifying global agent configuration files (like ~/.claude/CLAUDE.md) to load the skill in every future session. It also utilizes a background daemon process that stays active across individual command runs.\n- [DATA_EXFILTRATION]: The skill facilitates extraction of data from authenticated sessions (e.g., X/Twitter and LinkedIn). While the provided examples use restricted temporary files, the capability to read and process session-protected data could be exploited for exfiltration.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 29, 2026, 11:59 AM