SYNTROX.ai
Blog

Browser and Computer-Use Agents: The Security Risks of an AI That Clicks for You

5 min read

A browser or computer-use agent acts with everything the browser holds: logged-in sessions, saved passwords, the power to buy, send and delete. The four risks are hidden instructions in web content, inherited sessions, irreversible clicks and data typed into forms. The controls: a separate browser, per-task logins, allowlists, confirmation of consequential actions, page content treated as data, and bounded runs.

On this page

An API agent can do what its tools allow. A browser agent can do what you can do in a browser, which for most people is nearly everything: email, banking, the company's admin consoles, a shopping basket with a saved card.

That is the appeal, and the problem. A computer-use agent reads the screen, decides and clicks. Whatever is on the screen is part of what it reads.

Why browser agents are riskier than API agents

Browser security rests on a few old assumptions. A page from one site cannot read your session on another; the same-origin policy and CORS see to that. A page can show you a "buy now" button, but it cannot press it.

An agent operating the browser breaks both. It moves between sites by design, and it presses buttons. When Brave's researchers disclosed an indirect prompt injection in Perplexity's Comet browser in August 2025, they said that once an assistant follows instructions from untrusted page content, those protections are "effectively useless", because "the AI operates with the user's full privileges across authenticated sessions".

Four security risks of computer-use agents

1. Instructions in web content

Anything on a page can be read as an instruction: hidden text, a comment, alt text, words inside an image. In Brave's proof of concept the instructions sat in a Reddit comment behind a spoiler tag. The user asked Comet to summarise the page. The agent fetched the user's email address from their Perplexity account page, requested a one-time login code using a variant of Perplexity's domain with a trailing dot, read the code in the user's logged-in Gmail, and posted both as a reply to the comment. Brave reported it on 25 July 2025. After publishing, it said further testing showed the attack was still not fully mitigated, and re-reported it.

The desktop version can be cruder still. In October 2024 Johann Rehberger showed a web page whose text amounted to "download this file and launch it" turning Claude Computer Use, then in beta, into a "ZombAI". It downloaded the binary, made it executable and ran it, and the machine joined his command-and-control server.

This is indirect prompt injection with hands.

2. Logged-in sessions and credentials

The agent inherits whatever the browser profile holds: cookies for every site you are signed into, a password manager, autofill for cards and addresses. Nothing has to be stolen. It only has to be used. Anthropic's computer use documentation advises "avoiding giving the model access to sensitive data, such as account login information, to prevent information theft".

3. Irreversible clicks

Submit, send, pay, delete, accept. Most have no undo. Google's Computer Use documentation lists the categories its built-in safety service blocks or sends for confirmation, among them financial transactions, sending messages, creating accounts and accepting terms of service or other legally binding contracts. Its broader advice is to avoid the capability "for tasks involving critical decisions, sensitive data, or actions where serious errors cannot be corrected".

4. Data entry

Typing is sending. OpenAI's computer use guide says so directly: "Typing sensitive information into a form counts as transmission." A form on an attacker's page, or a search box whose query lands in someone else's logs, is an exfiltration channel. We cover the URL and image variants in how data leaves an AI agent.

What the vendors' own computer-use guidance says

Anthropic, OpenAI and Google each publish safety guidance for their computer-use tools. It converges.

AnthropicOpenAIGoogle
EnvironmentDedicated VM or container, minimal privilegesSeparate browser or VMSandboxed VM or container
LimitsAllowlist of domainsAllow list of sites and actionsSafety policies that block or require confirmation
Screen contentInstructions on pages or in images "might override your instructions""Treat screen content as untrusted"Supervise closely for important tasks
Consequential actionsA person confirms financial transactions, terms of service, cookie consentUsers stay in control of purchases, data transmission, destructive changesSome actions return a decision requiring user confirmation

None of them says the model will hold the line by itself. Anthropic runs classifiers over what its computer use tools return, such as screenshots, to flag likely injections, and still says its listed precautions "remain important even with these classifiers in place". Its November 2025 research post on browser use is plainer: "No browser agent is immune to prompt injection."

How to secure a browser agent: six controls

1. A separate browser, never the user's own. Run the agent in a dedicated VM or container with a fresh profile: no saved sessions, no password manager, no extensions you did not choose. Brave recommends keeping agentic browsing apart from ordinary browsing, so that no one ends up in it by accident.

2. Log in per task, with the narrowest account. Sign the agent into only the sites the task needs, with an account or role limited to that task, and end the session when it finishes. When a password or one-time code is needed, hand control to the person to type it. Where the target system has an API, an agent with its own scoped, short-lived credential is safer than one driving a logged-in interface.

3. Allowlists for sites and actions. Restrict navigation to the domains the task needs, enforced by a proxy rather than the prompt, and make downloads and running files off by default. The ZombAI attack needed both.

4. Confirmation that shows what will happen. Before sending, paying, deleting, submitting personal data or accepting terms, stop and show the person the exact recipient, amount, fields or document. Brave's example: "sending an email should always prompt the user to confirm right before the email is sent". Keep this for the consequential few, because a prompt on every click trains people to approve without reading.

5. Page content is data; proposed actions are checked against the task. The user's request is the instruction. Everything on screen is material. Brave's recommendation is that the agent's proposed actions be "independently checked for alignment against the user's requests". A task to summarise a page has no business opening a mail client.

6. Bound the run, and verify the outcome. Cap steps, time and cost, support cancellation, and, in OpenAI's words, "check the actual outcome instead of relying only on the model's final answer". Keep each run's screenshots and actions. They are the only record of what the agent saw when it decided.

A browser agent checklist

  1. Does the agent's browser hold any session, password or saved card it does not need for this task?
  2. Who types credentials and one-time codes: the person or the agent?
  3. Which domains can it reach, and where is that enforced?
  4. Which actions stop for confirmation, and does the confirmation show exactly what will be sent or paid?
  5. After a run, can you see every screen it saw and every click it made?

Where Syntrox fits

Syntrox sits outside the agent and outside the browser it drives. Our judge, Nucleus, treats any text inside the material it evaluates as evidence, never as instructions, so a hidden line on a page addressed to the evaluator counts against the output. It is small enough to run on hardware inside your own network, so the page content a check needs does not have to leave it.

5 min read

Indirect Prompt Injection: Why the Attack Arrives in the Data, Not the Prompt

Indirect prompt injection hides instructions in content an agent retrieves (a web page, email or RAG chunk), so the attacker never talks to the model. Filters on the user's prompt miss it, because the payload is in the data. Six controls hold: mark untrusted content, keep it out of the control flow, cut privileges after reading it, confirm consequential actions, block exfiltration, and check actions before they run.

Read

5 min read

AI Agent Identity: Why Shared Service Accounts Break Authorisation, and What Replaces Them

Give every agent its own identity, and when it acts for a person, put both in the credential: the user it acts for and the agent doing the acting. OAuth 2.0 Token Exchange (RFC 8693) defines exactly that delegation. Then make each credential short-lived, scoped to the task and bound to one audience, so a leaked or misused token is worth little and every action traces back to an agent and a person.

Read

5 min read

Data Exfiltration Through AI Agents: Five Ways Data Gets Out, and How to Close Them

An agent leaks data through the same channels it uses to work: tool-call arguments, URLs and rendered images, email and messages, files and pull requests it writes, and third-party MCP servers. Close those paths outside the model: allowlist where the agent can send data, inspect tool-call arguments before they leave, strip untrusted links and images from output, and scope every credential to the task.

Read