Skip to main content
Use this guide when your application lets a model browse pages, inspect a DOM, search the web, or take browser-side actions. Browser agents are useful because they can work with live content. They are risky for the same reason.

Assumptions

  • Your workflow can browse or inspect pages that contain untrusted text.
  • The application decides which browser actions are available.
  • Anthale can evaluate the content path before page text or browser results are reused.

Secure the browser workflow

1

Treat every page as untrusted content

Assume page text, rendered labels, metadata, and hidden instructions are untrusted until the next boundary check says otherwise. A familiar-looking interface is not a trust signal.
2

Separate reading from acting

Keep page inspection and navigation separate from actions such as form submission, purchase, deletion, or external messaging. A model should not turn a page instruction directly into a side effect.
3

Evaluate browser output before reuse

Run Anthale before browser results, copied snippets, extracted text, or page summaries are fed back into the model or shown to a user.
4

Constrain destinations and credentials

Use allowlisted domains, isolated sessions, and the smallest possible credential scope. Do not let one workflow browse broadly with the same privileges it uses for sensitive actions.
5

Require approval for high-impact browser actions

Put confirmation or approval in front of submits, sends, purchases, destructive changes, or actions that affect an external account.
6

Log page provenance and action decisions

Record which URL, domain, page source, or browser tool produced the content, and log whether the model only read, proposed, or actually executed an action.
Browser pages often mix useful content with attacker-controlled instructions. If the agent can both read and act in the same step without validation, the page becomes part of the control surface.

What this looks like in code

The model can read the sanitized browser content, but the application still decides whether the next browser action is allowed and whether it needs approval.

Common failure patterns

  • The model reads page text and immediately executes the next browser action without a validation step.
  • Allowed browsing domains are broader than the workflow requires.
  • Browser credentials have more privilege than the action path needs.
  • Teams log the final action but not the page or domain that shaped it.

Next steps

Continue with Gate Tool Actions, Validate Model Output, or Red-Team Agent Workflows.
Last modified on April 12, 2026