agent-browser
Pass
Audited by Gen Agent Trust Hub on Mar 16, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it extracts and processes content from external websites.
- Ingestion points:
agent-browser snapshot,agent-browser get text, and other data extraction commands. - Boundary markers: Supported via the
AGENT_BROWSER_CONTENT_BOUNDARIESenvironment variable, which wraps tool output in unique nonces. - Capability inventory: Shell command execution via
Bash, file system access (--allow-file-access,state save), and network navigation. - Sanitization: Content is delimited with nonces when boundaries are enabled.
- [DATA_EXFILTRATION]: The tool can access local files using
file:///URLs if the--allow-file-accessflag is provided. It also includes commands to read the system clipboard (agent-browser clipboard read). Session state files saved to disk contain session tokens in plaintext by default, though encryption is supported viaAGENT_BROWSER_ENCRYPTION_KEY. - [REMOTE_CODE_EXECUTION]: The
agent-browser evalcommand allows the execution of arbitrary JavaScript within the browser context, providing a mechanism for dynamic code execution. - [COMMAND_EXECUTION]: The skill allows the agent to execute the
agent-browserCLI to automate browser tasks. - [EXTERNAL_DOWNLOADS]: The instructions involve installing the
agent-browserpackage and downloading the Chromium browser engine.
Audit Metadata