in one lineAn AI browser agent is when an assistant reads web pages and then acts on them, clicking, typing and submitting, inside sessions where a person is already signed in.
the numbers
- 11.2%of prompt injection attacks still succeeded against a browser agent in its maker's own testsAnthropic, Piloting Claude for Chrome, 2025
1what it is
An AI browser agent reads web pages and then clicks, types and submits on them. Some run inside the browser people already use, as a sidebar or an extension. Others run a browser of their own on the vendor's computer.
The first kind works inside every tab where the person is signed in. OpenAI's documentation for its ChatGPT browser extension says it can read or act on sites where the user is already signed in, and names LinkedIn, Salesforce and Gmail as examples. Google's help page for auto browse in Gemini in Chrome says it has access to the same sites the user does, including the ones they are signed into. No new account or grant appears in your identity provider. The agent works through the person's own sessions.
Extensions that only read pages are covered in the AI browser extensions explainer; the full research and settings are in the browser agent guide.
2why it accumulates
It arrives one install at a time. A new browser or an extension is a decision someone makes on their own laptop. OpenAI's page says the Chrome permission prompt for its extension can include reading and changing all your data on all websites.
Permissions widen with each click. The ChatGPT extension asks before each new site, and one of its answers is to allow all sites, which OpenAI's page marks as elevated risk. Perplexity's help centre describes an "Always Allow" option in Comet that lets it act without asking for confirmation. Each choice outlasts the task.
Sessions outlive the task. OpenAI's help article on the cloud browser in ChatGPT Work says a sign-in there persists for future tasks until it expires, and that a task can keep running after the user closes their computer. When OpenAI retired its Atlas browser, its help article said deleting ChatGPT chats does not delete Atlas cookies, and warned against assuming that uninstalling Atlas removes that data.
The controls follow the company account. Microsoft's reference for the Edge policy that governs browsing with Copilot says it does not apply to a profile signed in with a personal Microsoft account. Perplexity's help centre describes its restrictions as settings for an organisation's members. An agent on a personal account, with a work tab open next to it, sits outside both.
3what it costs you
Text from every page, email and invite the agent reads reaches the model alongside the user's request, so whoever wrote it can steer the agent, as explained in the prompt injection explainer.
- Brave, 20 August 2025. Brave, which makes a competing browser, showed instructions hidden in a Reddit comment leading Perplexity's Comet, asked only to summarise the page, to fetch the user's email address, trigger a one-time login code, read it from the user's signed-in Gmail tab and post both back to Reddit. A later update on Brave's post says Perplexity had still not fully mitigated that kind of attack.
- OpenAI, 22 December 2025, about its own agent. In OpenAI's own automated red-teaming, an email planted in the inbox told the agent to send a resignation letter to the user's CEO. When the user asked for an out-of-office reply, the agent read that email and followed it. OpenAI wrote that prompt injection is "unlikely to ever be fully 'solved'".
- Anthropic, August 2025, about its own product. Anthropic's pre-launch tests of Claude in Chrome, 123 test cases in autonomous mode, found a 23.6% attack success rate without its safety mitigations and 11.2% with them. That is a vendor measuring itself before release.
A steered agent acts as the person: a message sent, a form submitted, a file shared. OpenAI's security page for ChatGPT Work says OpenAI keeps no separate complete record of Chrome actions taken through its extension. Because the agent uses the person's session, the mail or CRM log can show those actions under the person's name.
what it looks like when it goes wrong
Three ways this tends to go, each one taken from a real case.
- it happenedEchoLeak / Microsoft 365 Copilot, 2025howAn email with hidden instructions reaches an assistant that reads every inbox.what leavesanything the assistant can seewho ends up with itthe stranger who sent the email
- it happenedMCP tool poisoning, 2025howA tool description tells the agent to send files somewhere else.what leavesfiles the agent can openwho ends up with itwhoever wrote the tool
- it happenedReplit agent database deletion, 2025howAn agent with production access ignores the instruction to change nothing.what leavesthe live database, deletedwho ends up with ityour customers, who notice first
4what works
- Find them. Ask which AI browsers, sidebars and extensions people use for work, and read your managed browsers' extension list, as in the browser extensions guide.
- Set the browser policy explicitly. Google's admin help says auto browse in Chrome is controlled by
GeminiActOnWebSettingswith allowed and blocked URL lists, and is currently off by default for managed users. Microsoft's documentation says browsing with Copilot in Edge works only on domains on an admin allow list, with the block list taking precedence; it is in preview, needs a Microsoft 365 Copilot licence and is open only to tenants outside the European Economic Area. Put mail, banking, payroll and admin consoles on the block list. - Switch off the always-allow options. Perplexity's help centre says Comet admins can turn off "Always Allow" and set domains to read only or no access, and that domain settings override the global switch. OpenAI documents
features.browser_use_external = falseto turn off agent browsing through extensions in ChatGPT Work. - Keep agent work away from signed-in sessions. OpenAI's December 2025 post recommends limiting logged-in access when a task does not need it. A separate browser profile with no mail, finance or admin sessions signed in does that.
- Add agent sessions to the leaver checklist. OpenAI's help article shows where to clear the cloud browser's cookies, and the token guide explains why sessions should also be ended at the sites that issued them.
Which of your people has an agent that can open the finance tab, and who would notice if it did?
on the record
your turn
what you can do about it
this week
Set the AI agent policy in your managed browsers explicitly, with mail, banking, payroll and admin consoles on the block list.
In Chrome that is GeminiActOnWebSettings with its URL lists; in Edge it is AllowBrowsingWithCopilot with an allow list. Then ask staff which AI browsers and extensions they use outside those browsers.
go deeper
if you're not sure where to start
Run the offboarding checklist on one leaver. It covers every account and token to close when someone leaves.
open the checklist →
