browser
Warn
Audited by Gen Agent Trust Hub on Oct 20, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to navigate and interact with arbitrary web pages, which are untrusted sources of data. Malicious content on these pages could attempt to subvert the agent's instructions.
- Ingestion points: Web content is ingested via
superset browser navigate,eval,console, andcdpcommands, as well asbrowser-useinteractions. - Boundary markers: No specific delimiters or "ignore" instructions are used when processing content from the web pages.
- Capability inventory: The agent has high-privilege capabilities including shell access (
Bash), network operations (curl), file writes (screenshot --out), and remote publishing (superset pages publish). - Sanitization: There is no evidence of sanitization or filtering of data retrieved from web pages before it is processed or used in subsequent commands.
- [DATA_EXFILTRATION]: Several commands facilitate the movement of sensitive data.
- The
superset browser import-logincommand copies authentication cookies from local browsers (Chrome, Arc, Edge, Brave, etc.) into the agent's environment. - The
superset pages publishcommand uploads local files (screenshots) to a remote Superset workspace, which may inadvertently include sensitive session data or personal information visible in the browser. - The
superset browser cdpcommand generates a WebSocket URL that includes an embedded authentication token. - [EXTERNAL_DOWNLOADS]: The skill instructs the agent to install the
browser-usepackage usinguv tool install browser-useand may suggest usinguvxfor ephemeral execution of external code. - [DYNAMIC_EXECUTION]: The skill makes extensive use of dynamic code execution at runtime.
superset browser evaland CDPRuntime.evaluateare used to execute arbitrary JavaScript within the browser context.browser-useexecutes Python code provided via shell heredocs to perform complex automation tasks.- [PRIVILEGE_ESCALATION]: The
import-loginfunctionality on macOS specifically reads the browser's Keychain key, which is a protected system secret. While it requires a system-level user prompt, it represents an escalation of access to the user's saved credentials.
Audit Metadata