Přeskočit na obsah

AI that acts in your browser, and what to check before you trust it

AI browser agents can click, type and fill forms for you. See what they do well today, the risks, and the checks to make before you let one act.

Tým nerixuZveřejněno 5 min čtení

V tomto návodu

An AI browser agent is an assistant that does the clicking for you: you describe a task, and it opens pages, types, selects and fills forms until the task is done or it needs you. That works well today for small, checkable jobs. It is risky when the agent can send, pay or delete without asking, or when it follows instructions it finds inside a web page. This guide covers what agentic browsing can do now and the checks to make before you trust one.

What agents can do today

The big browser and AI companies all ship something here. Google says its auto browse feature in Chrome is in preview for Google AI Pro and Ultra subscribers, launched in the U.S., and pauses before sensitive steps such as purchases or posts. Anthropic's Claude in Chrome clicks, types and fills forms across tabs, and is on paid plans. Perplexity's Comet is a free browser with an assistant that does tasks; its help centre lists the browser agent's queries as a paid feature, with monthly limits by plan. Prices and availability change, so read each vendor's own page.

Tasks they handle reasonably well:

  • Collecting facts from several pages into one list or table.
  • Filling long forms from details you already hold, for you to review.
  • Repeating steps such as searching a site with a set of filters.
  • Preparing options for a trip, a quote or a supplier, without buying.

Tasks to keep for yourself, for now: anything with money, legal commitment, passwords, medical or banking sites, and anything you could not undo.

The risks, in plain words

  1. It does the wrong thing, quickly. A misread button or the wrong date can send a form that cannot be recalled.
  2. It is tricked by a page. Text on a web page can be written to look like an instruction. This is called prompt injection, and no agent is fully immune.
  3. It sees more than it should. An agent working in your logged in session can read what you can read.
  4. You cannot see what happened. If there is no record of the steps, you cannot learn or correct.

Seven checks before you let an agent act

  1. Does it show a plan first? You should see the goal, the steps and the sites it will use, before it starts.
  2. Does it stay on those sites? An agent that wanders to other sites is a risk.
  3. Does it stop before it sends, pays or deletes? Look for a clear approval card at every irreversible step, not one blanket yes at the start.
  4. Is Stop always on screen, and can you take over at any moment?
  5. What does it never touch? Password, card and one time code fields should be off limits, and CAPTCHAs should be left to you.
  6. Does it ignore instructions found on a page? Page text is content, not a command.
  7. What is sent where? Find out whether page text and your files go to a server, who runs it, how long it is kept and whether it trains models.

A tool that cannot answer these in plain words is one to test on a page you do not care about, or not at all.

How nerix handles it

We build nerix, so judge this section accordingly. nerix for Chrome has an Act mode in the side panel. For a task with several steps on a website, it shows a plan first: the goal, the steps and the sites it will use, and only those sites are used. It stops and asks before it sends, submits, pays, signs, deletes, shares personal data or downloads a file. Stop is always on screen, and you can take over at any moment.

The limits are written down, not left to a prompt. According to the extension's store notes, it never types into password, card, security code or one time code fields, does not solve CAPTCHAs and says "You do this part" instead. It never buys anything, never creates accounts for you and never submits a form without you. It does not follow instructions it finds written inside a web page, and the protected field rules are enforced in the extension's code. A workspace admin can set which sites it may work on and which tools it may use.

Here is what a task looks like. You type "Fill this form for a large pizza and stop before sending". nerix shows the plan, you approve it, it fills the fields and stops with a message that it needs your OK before the order is submitted. You press the button yourself, or choose not to.

On privacy: a page is read only when you ask or press a suggestion, on a site you allowed, and nothing is read in the background. Your question and the page text go to your nerix workspace to be answered, not to train models. When it fills a form, the AI sees the names of the fields, never the values. Read more on security.

Honest limits: nerix for Chrome is new, works on Chrome on a computer only, and its store listing is coming, so check the extension page for the current state. Until then you can try the chat and its tools in the app: chat with sources. AI can be wrong, so check what it filled and what it did.

A safe way to start

  • Pick a low risk task and a site you know.
  • Allow one site, not all of them.
  • Read the plan and remove any step you do not understand.
  • Watch the first run from start to the first approval card.
  • Look at the record of what it did afterwards, if the tool keeps one.
  • Widen slowly. Add one new kind of task at a time.

Should you use one at work

For a small team, an agent can save time on repeat forms and research. The rules to set are simple: which sites, which tools, who can approve, and what is kept in a log. Choose a tool whose admin controls match those rules, and keep humans on the steps that move money or send messages to customers. If you are comparing extensions, our guide to the best AI Chrome extensions for work lists what each does on a page.

The short version

AI that acts in your browser is useful for small, checkable tasks, and risky when it can send, pay or delete without asking. Before you trust one, ask for a plan, a stop button, an approval at each irreversible step, a ban on password and card fields, and a clear answer on where your data goes. nerix for Chrome is built around those, and you should test any agent on a low risk task first.

Na co se lidé ptají

What is an AI browser agent?
It is an assistant that does not only answer questions but also acts on websites for you: it opens pages, clicks, types and fills forms to finish a task you describe. The safe ones show a plan and stop for your approval at the risky steps.
Is agentic browsing safe?
It is safe enough for low risk tasks when you can see what it plans to do, stop it at any moment and approve each step that sends, pays or deletes. It is not safe to leave an agent alone on banking, health or sign in pages.
What is prompt injection?
It is text hidden in a web page, an email or a document that tells the AI to do something you did not ask for, such as sending data elsewhere. A safer agent treats page text as content to read, never as instructions to follow.
Which tasks should I try first?
Start with ones where a mistake is cheap, such as filling a form you review before sending, gathering details from several pages, or booking nothing but preparing the options. Add harder tasks once you trust how it stops.

Sdílet tento návod

V nerixu

Vyzkoušejte to na vlastní práci.

nerix napíše první verzi, odpovídá z vašich dokumentů se stránkou, ze které každá věta pochází, a drží práci týmu na jednom místě. Začněte s bezplatným tarifem.

Čtěte dál