TL;DR: An agentic browser lets an AI model navigate, read and act on websites for the user, while the user may be logged in to everything. Because every page is input to the model, any page can try to give it orders: Google calls indirect prompt injection "the primary new threat facing all agentic browsers", and OpenAI says it is "unlikely to ever be fully 'solved'". Browser makers answer with layers: classifiers on page content, a separate model that checks each action against the user's task, limits on which sites the agent may act on, and confirmation before purchases, posts and sign-ins. OpenAI's decision to retire its Atlas browser in August 2026 shows a second lesson: a browser is a long-term security commitment, not a feature.
A short history, 2025 to 2026
| Product | What happened | Source |
|---|---|---|
| Perplexity Comet | Limited release 9 July 2025; free to everyone from 2 October 2025, with background assistants | Perplexity announcement |
| Microsoft Edge Copilot Mode | Opt-in and labelled experimental from 28 July 2025; Copilot Actions in a limited US preview from October 2025 | Microsoft Edge blog |
| OpenAI ChatGPT Atlas | Launched 21 October 2025 on macOS with agent mode in preview; deprecated, with browser agent features moved into ChatGPT and Codex, and stopped working on 9 August 2026 | OpenAI launch post and help centre |
| Google Chrome auto browse | Preview for Google AI Pro and Ultra subscribers in the US from 28 January 2026, built on Gemini 3 | Google blog |
OpenAI gave its reason for retiring Atlas plainly: browsers "require ongoing security maintenance", and it did not want users left on a discontinued browser that stops receiving security updates. Building a browser means owning patch cadence, extension policy and compatibility for years. Adding agent capabilities to an existing browser or app avoids that burden.
Why the browser is the hardest place for an agent
A browser agent combines the three things that make prompt injection dangerous: it reads untrusted content (any web page), it has access to private data (the user's logged-in sessions), and it can act externally (send, buy, post). The prompt injection concept covers the general mechanism; browsers make it concrete.
Brave's security team demonstrated it in August 2025. A comment hidden behind a spoiler tag on a Reddit page told Comet, when the user asked for a summary, to find the user's email address, request a one-time password, read the code from Gmail and post both in a reply. Brave reported a partial fix within two days and later wrote that the issue was not fully mitigated. In October 2025 the same team showed instructions hidden in near-invisible text inside screenshots, and concluded the problem is "a systemic challenge facing the entire category". OpenAI's December 2025 post on hardening Atlas described an injected email that made the agent send a resignation letter to the user's CEO.
The layered defence
Follow the flow, one stage at a time.
Illustrated sequence; animation speed does not represent processing time.Read diagram description
The user gives the browser agent a task. The agent reads pages, which are untrusted input. A classifier screens page content for injected instructions. The planner proposes the next action. A separate critic checks the action against the user's task using only its metadata. An origin check decides whether the site is read-only or allows actions. Sensitive actions such as purchases, posts and sign-ins wait for the user's confirmation. Then the agent acts and reads the next page.
Google published the most detailed design, for agentic features in Chrome (December 2025). Its main layers:
- A prompt-injection classifier that runs in parallel with the planning model.
- A User Alignment Critic, a separate model that sees only metadata about each proposed action and judges whether it serves the user's task. Because it never reads the page, page text cannot argue with it.
- Agent Origin Sets, separating sites the agent may only read from sites where it may click and type.
- Confirmations before navigating to certain sensitive sites and before signing in through the password manager; Chrome auto browse also pauses before purchases and social posts.
OpenAI's Atlas agent could not run code in the browser, download files, install extensions or reach the computer's file system; it paused on sensitive sites such as banks; and it offered a logged-out mode. OpenAI also trained an automated attacker with reinforcement learning to find new injections and retrained the agent against them.
Two product lessons follow. The defences reduce risk without removing it, so the scope of what the agent may do unattended has to match that residual risk. And the safest permission can be no login at all: an agent that browses logged out cannot spend the user's money or read their email.
If agents are visiting your site
Agents now arrive as customers. Cloudflare reported in July 2026 that more than half of internet traffic is non-human (company-reported). The question for a site owner shifts from "block bots" to "which agents do we trust, and for what?". Cloudflare's Web Bot Auth lets an agent sign its requests cryptographically, and an IETF working group chartered in October 2025 is standardising the method; as of November 2026 the specification is still a draft. A signature identifies the agent, not the person behind it, so it pairs with the user-level controls in delegated access for agents and the commerce layers in agentic commerce protocols.
In the interview
Expect "should we build an AI browser?", "design a browser agent that books travel" or "our site is getting agent traffic; what do we do?". A weak answer demos a clever automation. A Hire answer adds confirmations. A Strong Hire answer names indirect prompt injection as the core risk, layers content screening, an action critic, origin limits and confirmations, prefers logged-out or narrowly scoped sessions, prices the maintenance of owning a browser, and on the site side separates identifying an agent from authorising what it may do.
Practise with which capabilities may a browser agent combine and should we build a browser agent or use APIs.
Next: computer-use agents and prompt injection.