agent-browser
Pass
Audited by Gen Agent Trust Hub on Mar 3, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- Dynamic JavaScript Execution: The skill includes an
evalcommand that allows the execution of arbitrary JavaScript within the browser context. While this is a standard requirement for complex browser automation, it allows the agent to interact with the page's internal state and should be monitored for unintended actions. - Evidence found in
references/commands.mdandSKILL.md. - Local File System Access: The skill supports the
--allow-file-accessflag, enabling the browser to open local files (e.g.,file:///path/to/doc.pdf). This allows the agent to process local documents but also creates a surface for data exposure if the agent is directed to a malicious site while local files are open. - Evidence found in
SKILL.mdunder 'Local Files'. - Session and State Persistence: The
state saveandauth savecommands persist session data, including cookies and localStorage, to JSON files. This is a functional feature for maintaining logins across sessions, but these files contain sensitive session tokens and credentials that should be protected from unauthorized access. - Evidence found in
templates/authenticated-session.shandreferences/session-management.md. - Potential for Indirect Prompt Injection: The skill ingests untrusted data from the web via snapshots and text extraction. This introduces a risk where malicious instructions on a webpage could attempt to influence the agent's behavior.
- Ingestion points: Web content is ingested via
agent-browser snapshotandagent-browser get textintemplates/capture-workflow.shandSKILL.md. - Boundary markers: The skill provides an opt-in
--content-boundariesflag (detailed inSKILL.md) to help the agent distinguish between tool output and page content. - Capability inventory: The skill possesses powerful capabilities including
eval(JS execution), file-write (screenshots, PDFs), and navigation to arbitrary URLs. - Sanitization: Content boundaries are used to delimit external content, reducing the risk of the agent obeying instructions embedded in web pages.
Audit Metadata