test-browser
Pass
Audited by Gen Agent Trust Hub on Feb 17, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [EXTERNAL_DOWNLOADS] (LOW): The skill installs the
agent-browserpackage globally via npm and executesagent-browser installto download Chromium (~160MB). - Evidence:
npm install -g agent-browserandagent-browser installin SKILL.md. - Context: While downloading executables is a high-risk activity, this package is attributed to Vercel, which is a trusted organization in the security policy, downgrading the severity to LOW.
- [COMMAND_EXECUTION] (SAFE): The skill executes various system commands including
git,gh, andnpm. - Evidence: Uses
git diff,gh pr view, andnpm install. - Context: These commands are essential for the skill's primary purpose of identifying changed files in a PR/branch and testing them, following the intended workflow.
- [DATA_EXFILTRATION] (SAFE): The skill interacts with
localhost:3000and captures snapshots/screenshots. - Evidence:
agent-browser open http://localhost:3000andagent-browser snapshot -i. - Context: All network operations are directed at the local development server (whitelisted) or via the GitHub CLI for PR metadata. No data is sent to unknown external domains.
- [INDIRECT_PROMPT_INJECTION] (LOW): The skill has a surface for indirect prompt injection as it ingests untrusted data from web pages.
- Ingestion points: Web page content and interactive element metadata retrieved via
agent-browser openandagent-browser snapshot -ifromlocalhost:3000. - Boundary markers: Absent. The instructions do not specify delimiters to separate page content from the agent's internal logic.
- Capability inventory: Significant capabilities including
npm install,gitandghcommand execution, and file writing (creating todo files). - Sanitization: Absent. The skill does not describe any validation or escaping of the content retrieved from the browser before the agent processes it.
- Risk: An attacker who can influence the content of the local web application (e.g., via a malicious PR) could embed instructions in the HTML designed to manipulate the agent into executing unintended commands.
Audit Metadata