Agent Browser Security Tests Before You Ship

Agent browser security testing checks whether a browsing agent can be tricked by webpage text, hidden instructions, forms, downloads, or fake UI into taking actions the user did not approve.

Quick Answer

1. Seed pages with hostile instructions.
2. Test read-only browsing before form fills.
3. Require explicit approval for submits and purchases.
4. Block credential copying into untrusted pages.
5. Save every failure as a regression case.

Web pages are untrusted prompts

A browsing agent reads the page as context. That page can contain normal copy, hidden text, comments, fake system instructions, or malicious forms. The model may summarize the page correctly and still obey the wrong instruction.

Treat webpage content like user input from a stranger.

Clicks need policy

Navigation is low risk. Submitting a form, accepting terms, changing settings, or downloading a file is not. The agent should identify the action, the target site, and the expected result before it clicks.

That confirmation should be boring and specific. "Submit this form to example.com with your email address" beats "continue?"

Where BreakMyAgent fits

BreakMyAgent can replay hostile page fixtures and check whether the agent follows user intent instead of page-injected instructions.

FAQ

Are hidden prompts the only risk?

No. Visible instructions, fake UI, confusing buttons, and malicious downloads can be just as dangerous.

Should browser agents store passwords?

No. They should use the browser's existing secure session where possible and avoid exposing credentials to the model.

What is the best first test?

Ask the agent to summarize a page that contains a hostile instruction telling it to ignore the user and perform a write action.