browser-use
Fail
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: HIGHDYNAMIC_EXECUTIONDATA_EXFILTRATIONCREDENTIALS_UNSAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [DYNAMIC_EXECUTION]: The skill provides a high-risk interface for arbitrary code execution via the
browser-use pythoncommand. - Evidence in
references/cdp-python.mddemonstrates that arbitrary Python statements can be executed in the host environment, with variables persisting across calls. - The
browser-use evalcommand allows for the execution of arbitrary JavaScript within the browser context, enabling the manipulation of web pages and access to browser-scoped data likelocalStorage. - [DATA_EXFILTRATION]: Multiple commands facilitate the extraction of sensitive authentication data.
- Commands such as
browser-use cookies get,browser-use cookies export, andbrowser-use get valueallow for retrieving session tokens and user inputs. - The
browser-use screenshotcommand can capture sensitive information displayed on the screen. - [CREDENTIALS_UNSAFE]: The skill manages and accesses highly sensitive credentials.
- It provides access to local Chrome profiles via the
--profileflag, which contains the user's saved passwords, cookies, and browsing history. - Commands like
browser-use cloud loginandbrowser-use cloud signupinvolve the handling and storage of API keys in local configuration files (~/.browser-use/config.json). - [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to interact with external, untrusted web content, creating a significant attack surface.
- Ingestion points:
browser-use open,browser-use state,browser-use get html, andbrowser-use get textfetch data from the live web into the agent's context. - Boundary markers: No explicit instructions are provided to the agent to treat fetched web content as untrusted or to ignore embedded instructions.
- Capability inventory: The agent has access to powerful capabilities including arbitrary command execution (CLI), Python/JS execution, and network operations.
- Sanitization: There is no evidence of sanitization or filtering of the content retrieved from web pages before it is processed by the agent.
- [COMMAND_EXECUTION]: The skill utilizes a broad range of CLI commands that interact with the operating system, including starting daemons, managing network tunnels, and interacting with local browser installations.
- [EXTERNAL_DOWNLOADS]: The
browser-use profile updatecommand initiates the download and installation of external binaries (profile-use) to the local system.
Recommendations
- AI detected serious security threats
Audit Metadata