vercel-agent-browser
Warn
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: MEDIUMMETADATA_POISONINGCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [METADATA_POISONING]: The skill is titled 'Vercel Agent Browser Skill', which may lead users to believe it is an official tool from Vercel (a trusted organization). However, the metadata indicates the author is 'reason-machines' and the description attributes it to 'ara.so'. This discrepancy in branding is misleading.
- [COMMAND_EXECUTION]: The installation instructions require running global npm/cargo/brew commands to install
agent-browser. Furthermore, the commandagent-browser installdownloads external binaries (Chrome for Testing), andagent-browser upgradeexecutes remote scripts to update the tool, which are significant execution surfaces. - [REMOTE_CODE_EXECUTION]: The skill relies on the download and execution of the
agent-browserpackage from public registries and subsequently downloads binary executables (Chrome) during the initialization phase, which represents a supply chain risk. - [INDIRECT_PROMPT_INJECTION]: The skill facilitates browser automation and content extraction, creating a large surface for indirect prompt injection. Ingestion points:
agent-browser open,snapshot,get text, andevalingest content from untrusted external URLs. Capability inventory: The tool can manage network requests, access local storage/cookies, take screenshots, and execute arbitrary JavaScript (eval). Sanitization: There is no mention of sanitizing or escaping the content retrieved from web pages before the agent processes it. - [DYNAMIC_EXECUTION]: The skill provides an
evalcommand that allows the execution of arbitrary JavaScript code within the browser context. This can be used to perform actions or extract data that the standard CLI commands do not support, but it also provides a mechanism for executing code derived from untrusted sources. - [CREDENTIALS_UNSAFE]: The documentation includes examples of handling sensitive credentials, such as setting session IDs, tokens, and passwords via
agent-browser cookies set,agent-browser storage local set, andagent-browser set credentials.
Audit Metadata