The browser has become the newest AI battleground. Perplexity's Comet, OpenAI's ChatGPT Atlas, Opera Neon, The Browser Company's Dia, and AI features built into Chrome and Edge all promise the same thing: an assistant that sees what you see and can act on your behalf. Since PDFs open in the browser all the time, it is worth knowing what these assistants do with them.
What an AI browser can do with a PDF
- Summarise or answer questions about the PDF in the current tab.
- Compare documents across several open tabs.
- Extract data, such as dates, amounts or names, into a table.
- Fill in web forms using information from a PDF, in agent mode.
- Download, rename or move files in some agent setups.
Where your document goes
This is the question that matters. When you ask an AI browser about a PDF, the text (and often page images) is usually sent to the provider's cloud model. Some browsers run small models on device for simple tasks, but most summarisation and agent actions are cloud-based.
That is fine for a public report. It deserves thought for a medical record, an employment contract, or anything under a confidentiality agreement. Check:
- Whether the browser's AI features are on by default.
- Whether your content may be used for model training, and how to opt out.
- Whether your organisation allows the browser at all. Many IT teams have blocked agentic browsers pending review.
The prompt injection problem
Agentic browsers read page content and then take actions. Security researchers have repeatedly shown that hidden instructions in web pages or documents can hijack an agent, for example to open another site or reveal information from a different tab. A PDF with invisible text is an easy vehicle for that. See hidden prompt injection in PDFs.
Practical rules:
- Do not run agent mode while logged in to banking, email or admin consoles you would not want an agent to touch.
- Watch and approve actions, especially anything that sends, buys or deletes.
- Be suspicious when a summary contains instructions or urges you to visit a link.
Opening sensitive PDFs safely
If you just need to read, sign or fill a sensitive PDF, you do not need AI at all. Keep those tasks in a tool that processes files locally:
- Use a regular browser profile with AI features off for confidential documents.
- Use a browser editor that does not upload files, such as Docento.app, for signing and form filling.
- Download and open in a desktop reader when in doubt.
When AI browsing is genuinely useful
Comparing three public vendor brochures, summarising a long public consultation, or pulling the key dates out of a published policy: these are great uses with little downside. The rule of thumb is simple. If you would happily paste the document into a public chatbot, an AI browser is fine.
Takeaway
AI browsers make PDFs easier to digest, and they also make it easy to send documents to the cloud without noticing. Know which mode you are in, keep agents away from sensitive sessions, and use local tools for confidential paperwork. For more, see are online PDF editors safe? and privacy in browser PDF editing.
