agent-browser
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill navigates to and extracts information from arbitrary external web pages via commands like
open,snapshot, andget. This content is processed by the agent to determine subsequent steps, which creates a surface for indirect prompt injection where malicious websites could influence the agent's behavior. - Ingestion points: Content retrieved from the web via tools described in
SKILL.md(e.g.,snapshot,get text,get html). - Boundary markers: None identified in the provided files to delineate external content or instruct the agent to ignore embedded commands.
- Capability inventory: The skill provides capabilities such as clicking, form filling, cookie management, and JavaScript execution (
eval), which could be leveraged in an injection scenario. - Sanitization: No sanitization or filtering logic was found for the web content before it enters the agent context.
- [DYNAMIC_EXECUTION]: The
evalcommand documented inSKILL.mdallows the agent to execute arbitrary JavaScript code within the context of the browser. This provides a direct mechanism for dynamic code execution based on agent-generated or page-sourced input. - [EXTERNAL_DOWNLOADS]: The
check-update.shscript facilitates the download of binary updates from thevercel-labs/agent-browserrepository on GitHub. - [COMMAND_EXECUTION]: The skill includes a shell wrapper (
bin/agent-browser) that executes native platform binaries. It also contains development and utility scripts for cross-platform builds using Docker and for synchronizing skill files between local directories.
Audit Metadata