in this guide
On 20 August 2025 Brave, which makes a competing browser, published a demonstration against Perplexity’s Comet. Someone left instructions hidden in a Reddit comment. A user opened the thread and asked Comet’s assistant to summarise the page. The assistant read the hidden text as part of its task, went to the user’s Perplexity account settings for their email address, started a login that sent a one-time code, opened the user’s Gmail tab, which was already signed in, read the code, and posted the email and code back to Reddit as a reply.
Nobody typed a password. Nobody granted a new permission. Every step used a session the user already had open.
That is the mechanism. A browser agent is an AI that reads pages and then clicks, types and submits on them. When it runs in your own browser, it works inside every tab where you are signed in: mail, CRM, payroll, the admin console. And it takes in text from every page it reads through the same channel as your request, so whoever wrote that text gets a say in what it does.
The agent guide asks which identity an agent runs under. A browser agent answers that question in the bluntest way: it runs as the person, in the person’s sessions, with nothing new to see in your identity provider. The coding agent guide covers the same pattern in a developer’s terminal. This guide is about the browser.
the comet demonstration, as brave described it
- 1hidden textInstructions placed in a Reddit comment.
- 2an ordinary requestThe user asks the assistant to summarise the thread.
- 3the page becomes the bossThe assistant treats the comment as part of its instructions.
- 4your sessions do the restAccount settings, then the signed-in Gmail tab, then a public reply with the code.
two places the agent can run
The products changed names between 2025 and 2026, and ChatGPT Atlas, OpenAI’s own browser, launched on 21 October 2025, per OpenAI’s launch post, and has since been retired, per OpenAI’s help centre. Underneath the names there are two designs, and they put your sessions in different places.
The agent inside your own browser. An extension or a built-in sidebar. OpenAI’s documentation for its browser extension says ChatGPT can “read or act on sites where you’re already signed in”, and names LinkedIn, Salesforce and Gmail as examples. When Chrome installs it, the permission prompt can include reading and changing all your data on all websites. Google’s help page for auto browse in Gemini in Chrome says it has access to the same sites you do, including those you are signed into, and can use Google Password Manager with your permission. Comet, Anthropic’s Claude in Chrome and Microsoft Edge’s “browsing with Copilot” also act inside the user’s own browser.
The agent in a browser of its own. OpenAI’s cloud browser in ChatGPT Work runs on a remote computer, per OpenAI’s help article. It keeps its own cookies and sessions and does not use the sign-ins, cookies or saved passwords in your own browser. When a site needs a login, you sign in through a form the model cannot see. After that, the article says, the authentication persists for future tasks until it expires, and a task can keep running after you close your computer.
where the signed-in sessions live
agent in your browser
extension or sidebar
- Every tab you are signed into, from the first task
- Sometimes the saved passwords, with a prompt
- Your browser policies can reach it
agent in a cloud browser
on the vendor's computer
- Signed in only where someone logged it in
- Those sessions stay until they expire or are cleared
- Keeps working with your laptop shut
- Outside your device management
The first design reaches more on day one. The second accumulates: each task can leave one more signed-in session behind, on a computer you do not manage.
what the research showed, in order
Most of the research below targeted Comet. The same weakness is named by the other vendors in their own documents. Each item comes from a company with a commercial interest: security vendors, a rival browser maker, or the product’s own maker.
- 20 August 2025, Brave. The Reddit demonstration above. Brave reported it on 25 July 2025; by Brave’s account Perplexity shipped a first fix two days later, Brave found it incomplete, and an update added to the post said Perplexity still had not fully mitigated this kind of attack.
- 20 August 2025, Guardio Labs. Asked to buy a smartwatch on a fake shop, Comet in some runs completed the purchase and filled in the saved address and card by itself; in other runs it refused or handed checkout back. Given a phishing email, it clicked the link and helped fill in the login.
- 4 October 2025, LayerX. A crafted link made Comet pull data from connected Gmail and Calendar and send it to the researchers’ server, encoded so that Perplexity’s checks missed it. LayerX says Perplexity replied that it could not identify any security impact.
- 21 October 2025, Brave. Faint text hidden in an image steered Comet when a user asked about a screenshot. In another browser, Fellou, simply opening a page was enough for its text to reach the model.
- 22 December 2025, OpenAI, about its own product. In a demonstration, OpenAI’s automated attacker planted an email in the user’s inbox. When the user asked the Atlas agent to draft an out-of-office reply, the agent read that email and sent a resignation letter to the CEO. OpenAI says that after a security update the agent detects this attempt, and that prompt injection is unlikely to ever be fully “solved”.
- March 2026, Zenity Labs. Text in a calendar invite hijacked Comet during a routine task; the agent read local files and sent them to an outside server. Help Net Security reported a second variant aimed at password-manager workflows. Zenity says Perplexity fixed the issue before disclosure.
Vendors publish their own numbers too. Anthropic’s own red-team tests of Claude in Chrome, published in August 2025, found a 23.6% attack success rate without its mitigations and 11.2% with them in autonomous mode, across 123 test cases. That is a vendor measuring itself, and it is still roughly one attack in nine.
The agent can reach what your sessions reach. Anyone who can put text in front of it can ask.
The same shape appeared in an assistant that reads mail: the EchoLeak case, where a crafted message could steer Microsoft 365 Copilot toward data its user could see.
what the vendors give an admin
Every major vendor now documents controls. They differ in what they cover and how they behave when left alone.
- Microsoft Edge. “Browsing with Copilot” is in preview, needs a Microsoft 365 Copilot licence, and is open only to tenants outside the European Economic Area, per Microsoft Learn. It works only on domains an admin puts on an allow list; with an empty list it is unavailable, and the block list always wins. It asks the user to handle sign-in steps and the final save or submit. The
AllowBrowsingWithCopilotpolicy turns it on or off for everyone who receives it. - Google Chrome. Gemini in Chrome needs both the Gemini app setting and the
GeminiSettingspolicy. Auto browse has its own policies:GeminiActOnWebSettingsplus allowed and blocked URL lists. Google’s admin help says auto browse is currently off by default for managed users, while the policy reference describes an unset value as allowed. Setting it explicitly removes the doubt. - OpenAI. The extension asks before each new site, with options up to “Allow for all sites”, which OpenAI’s page marks as elevated risk. Enterprise admins can switch off agent browsing through extensions with
features.browser_use_external = false. The cloud browser has a setting with three levels: always ask, auto approve, and always allow, which OpenAI says is not recommended. - Perplexity Comet. Admins of an Enterprise organisation can switch off the assistant’s control of the browser, remove “Always Allow” so every action needs approval, and set individual domains to browser control, read only or no access. Perplexity’s help page warns that domain settings override the global ones, so a domain left on browser control stays reachable after the global switch goes off.
Google’s December 2025 security post describes the defences behind the scenes: a second model that checks each planned action, limits on which sites the agent may act on, and confirmations before banking or medical sites. Google calls prompt injection an open challenge. Gartner went further. Its advisory “Cybersecurity Must Block AI Browsers for Now”, as reported by The Register on 8 December 2025, told security teams to block AI browsers where the AI service behind them fails a risk assessment.
where the controls stop
Each of those controls is attached to something specific, and it is worth knowing what.
They follow the company account or the managed browser profile. Comet’s restrictions apply to members who sign in with their Enterprise account. Microsoft’s policy page says the Edge policy does not apply to a profile signed in with a personal Microsoft account. Chrome’s agent policy applies at the browser profile level. Google’s admin page says people without a qualifying Workspace edition use Gemini under consumer terms. A personal AI browser, or a personal account in an extension, with a work tab open, sits outside all of it.
The record is partial. OpenAI’s security page for ChatGPT Work says OpenAI keeps no separate complete record of Chrome actions taken through the extension, and warns against assuming every action shows up in its Compliance API. Beyond that, the agent acts through your session, so in a mail or CRM log its actions can look like yours.
The sessions outlive the task. A cloud browser keeps its sign-ins until they expire or someone clears them in settings. When OpenAI retired Atlas, its help article told users that deleting chats does not delete Atlas cookies, and that uninstalling Atlas should not be assumed to delete that data either. When someone leaves, their agent’s browser is one more place a session can survive. The token guide explains why a session keeps working until someone ends it at the site that issued it.
Like the add-ons in the browser extensions guide, these agents arrive one install at a time. And like the connectors in the MCP servers guide, they are a decision someone made on a laptop. For the short explainer, see AI browser agents.
If someone hid one sentence for an agent in the next calendar invite your finance lead accepts, which signed-in tabs would that agent find open?
sources
- Brave: Agentic browser security, indirect prompt injection in Perplexity Comet (20 August 2025)
- Brave: Unseeable prompt injections in screenshots, more vulnerabilities in Comet and other AI browsers (21 October 2025)
- Guardio Labs: Scamlexity, we put agentic AI browsers to the test (20 August 2025)
- LayerX (now on Akamai's blog): How one click can turn Perplexity's Comet AI browser against you (4 October 2025)
- Zenity Labs: PerplexedBrowser, accepting a meeting or handing your local files to an attacker (5 March 2026)
- Help Net Security: The vulnerability that turns your AI agent against you (4 March 2026)
- OpenAI: Continuously hardening ChatGPT Atlas against prompt injection attacks (22 December 2025)
- OpenAI: Introducing ChatGPT agent (17 July 2025)
- OpenAI: Introducing ChatGPT Atlas (21 October 2025)
- OpenAI Help Center: ChatGPT Atlas retirement
- OpenAI Help Center: Using cloud browser in ChatGPT
- ChatGPT docs: Browser extension
- ChatGPT docs: ChatGPT Work local security
- Anthropic: Piloting Claude for Chrome (25 August 2025, updated)
- Google Chrome Help: Ask Gemini in Chrome to complete tasks for you with auto browse
- Chrome Enterprise and Education Help: Gemini in Chrome
- Chrome Enterprise policy list: GeminiActOnWebSettings
- Google: Architecting security for agentic capabilities in Chrome (8 December 2025)
- Microsoft Learn: Configure browsing with Copilot
- Microsoft Learn: Microsoft Edge policy AllowBrowsingWithCopilot
- Perplexity Help Center: Managing Comet Assistant permissions
- Perplexity Help Center: My organization has restricted Comet
- The Register: Block all AI browsers for the foreseeable future, Gartner (8 December 2025)
