Perplexity Comet security risks: is Comet safe for work?
Comet puts an AI assistant and an agent inside every tab your staff are signed in to. Whether that is safe depends on what the agent can reach and who manages it. Here is how to judge it, and what to do either way.
The AI Agent Blocklist: block Comet and every other AI browser and agent extension by risk grade
- Comet's macOS and Windows identifiers next to every other AI browser
- Agent extension IDs for Chrome and Edge policies
- Domain lists for firewalls and DNS filters, updated every day
Is Comet safe for work? The short answer
Comet is a mainstream product from a known company. The question for IT is not whether the vendor is trustworthy. It is whether an AI agent with your users' sessions fits your risk tolerance today.
Not a fit for most organisations. Nobody set its policies, its data terms are the user's own, and IT cannot see what the agent did.
Possible for a named group, if you deploy it through MDM, set its policies, limit the agent and accept that prompt injection is not solved.
Block it, as Gartner advised for AI browsers in December 2025. Review again when controls and research move on.
Gartner's advisory "Cybersecurity Must Block AI Browsers for Now" covers AI browsers as a category, not one product. The same questions apply to every AI browser your staff might install. See what the Gartner advice means.
What Comet can see and do on a work laptop
Every security question about Comet comes back to this chain. The assistant reads, the AI service works on what it read, and the agent acts with the user's access.
Sees
The open page, other tabs and the text or files the user shares with the assistant.
Sends
What it reads goes to Perplexity's AI service to be answered or planned.
Acts
The agent clicks, types, navigates and fills forms in sessions where the user is signed in.
Connects
Users can link services such as mail and calendar. Perplexity says Comet runs local MCP servers only with user permission.
The weak link is the page, not the browser. Anything the assistant reads can contain instructions. A web page, a shared document or an email can try to steer the agent. Our page on AI browser prompt injection explains how.
Comet compared with the browser you already manage
The risk is not that Comet has AI. Your managed browser may have AI features too. The difference is who controls the AI layer and whether an agent can act.
| Question | Your managed browser | Comet, consumer install | Comet Enterprise |
|---|---|---|---|
| Who sets the policies? | IT, through Group Policy, Intune or the Google Admin console | Nobody. The user's own settings apply | IT, through MDM and Chromium policies |
| Can IT switch the AI features off? | Yes, where the browser offers AI policies | No | Check the current admin console for Assistant and Agent controls |
| Can an agent act in signed-in sessions? | Only if you allow agent features or agent extensions | Yes, whenever the user asks it to | Depends on the controls you set |
| Which account and terms apply? | Your contracts with your browser and AI vendors | The user's personal account and consumer terms | Perplexity Enterprise accounts and enterprise terms |
| Does IT see it in inventory? | Yes | Only if you search for its identifiers | Yes, as a deployed app |
Agent extensions bring the same question into the browser you manage. If you block Comet but allow every extension, the agent comes back another way. Read are AI Chrome extensions safe.
Risk patterns that apply to every AI browser
Researchers showed in 2025 that AI browser assistants in general can be steered by what they read. These patterns belong to the whole category, not to one product, and makers describe prompt injection as an open problem for the industry.
- Page text
Text that people cannot see, in a page or a comment, can reach the assistant as if it were the user's request. Keep the agent away from mail and admin consoles so a steered task reaches little.
- Links
A crafted link can carry instructions inside the web address. Users who open links from outside are the exposed group, so link training and role-based blocks matter.
- Web address
Text after the # in a web address never reaches the server, so network tools do not see it, while the assistant can still read it. Network rules alone cannot catch this pattern.
- Connectors
Any agent that can start local tools or MCP servers reaches files and programs on the device. Treat every local connector as a security approval with a named owner.
None of these patterns needs malware on the device. They use the assistant's own access. That is why blocking or tightly managing AI browsers is a policy question, not a patching question. More in AI browser security risks.
Comet risk register for IT teams
Use this register in your risk assessment. Priority is our suggestion for an organisation that has not yet decided on AI browsers.
| Risk | How it happens | Who is exposed | Control | Priority |
|---|---|---|---|---|
| Company data reaches an AI service outside your contracts | The assistant reads open pages and tabs and sends them to be processed | Anyone with customer, patient, student or financial data on screen | Block consumer Comet; if allowed, Enterprise accounts only and staff told to close sensitive tabs | Act now |
| Agent follows instructions from a page | Hidden text in a page, comment, document or email steers the assistant | Every user, most of all those signed in to mail and admin consoles | Block the agent, or limit it to approved sites and keep it away from email | Act now |
| Crafted links trigger actions | Instructions ride inside a web address sent by mail or chat | Users who click links from outside | Block the browser for high risk roles; train users on links | Plan |
| Wrong actions in real accounts | The agent misreads a page and submits a wrong form, order or booking | Finance, procurement, HR, travel bookers | No agent use in finance and HR systems | Plan |
| Passwords and sessions in an unmanaged browser | Users save passwords and stay signed in inside a browser IT does not manage | All users of the browser | Approved password manager policy; remove profiles when you uninstall | Act now |
| Local access through MCP | Users approve local MCP servers that give the agent access to files or tools | Developers and power users | MCP allowlist and blocks on unapproved servers | Plan |
| Software outside your patch process | Per-user installs that update themselves and never appear in your packaging | The whole fleet | App control rules and inventory by identifier | Plan |
| Training and attestations done by a bot | Staff let the agent click through mandatory courses or forms | Compliance, HR, schools | Policy, and agent blocks on training platforms | Watch |
To score these for your own context, use our AI agent risk assessment. For local connectors, see MCP security risks.
Who carries the most risk
The same browser is a small risk for a marketing intern reading public pages and a large one for someone with admin rights. Start your rules with these groups.
Finance and procurement
Payments, purchase orders and supplier portals are exactly where a wrong agent action costs money.
Executives and assistants
Their mail and calendars are the richest target for data that leaves through an assistant.
IT and cloud admins
An agent in a tab signed in to a cloud console works with admin rights.
Developers
Source code, tokens and local tools sit one MCP approval away from the agent.
HR, legal and health
Records with special legal protection must not reach an unapproved AI service.
Schools
Students can hand coursework and tests to an agent. Staff hold student records.
Shared and kiosk devices need special care. An AI browser left signed in on a shared laptop gives the next user an agent that acts as the last one. Block AI browsers on shared devices even if you allow them elsewhere.
Contractors and personal devices. You cannot enforce app blocks on devices you do not manage. Cover them with network rules on your Wi-Fi, conditional access for company apps and a clear line in your AI policy.
Signs Comet is already in use
Most teams find it before they decide on it. Look in these places first.
- Software inventory lists Comet, or the macOS bundle ID
ai.perplexity.comet - Windows inventory shows the package
Perplexity.Comet - DNS logs show Perplexity traffic from devices with no approved use
- Users report a new default browser or ask for Comet by name
- Third-party app consents for Perplexity appear in your Google Workspace or Microsoft 365 admin center
- Help desk tickets about sign-in prompts in a browser you did not deploy
A wider sweep for AI tools is in our guide to shadow AI agents.
If you allow Comet: settings that lower the risk
Some teams decide to allow Comet for a group with a clear use case. These are the conditions we would set. If you cannot meet them, block it.
Deploy and sign in
- Use Comet Enterprise, launched in March 2026, and deploy it through MDM with the silent installer on Windows and macOS.
- Require sign-in with Perplexity Enterprise accounts. Block consumer installs for everyone else.
- Scope it to a named group with a business owner who reviews use each quarter.
- Read the current enterprise terms for data retention and training, and record the answers.
Configure and watch
- Set its Chromium policies: block extensions except those you approve, and restrict sites.
- Decide whether the Assistant and the Agent are allowed at all. Check the current admin console for these controls.
- Keep the agent away from email, finance systems and admin consoles.
- Tell users not to keep sensitive pages open while they use the assistant, and to review every action the agent proposes.
Managing the browser lowers the risk. It does not remove prompt injection. Keep the allowed group small and the block in place for everyone else. Steps for the block are on how to block Perplexity Comet.
What to tell staff about Comet
A block without a reason leads to workarounds. A short message, sent before the block, keeps tickets down and gives people a route to ask for an exception.
What the message should say
- AI browsers such as Comet are not approved on company devices, and why in one sentence: they let an AI act in your accounts.
- The date the block starts, and that existing copies will be removed.
- How to keep bookmarks before removal, and that saved passwords belong in the approved password manager.
- Which AI tools are approved, and how to request a new one.
What users should do if they used Comet
- Sign out of company accounts inside Comet before it is removed.
- Remove any company mail or calendar connections they made.
- Tell IT if the agent ever took an action they did not expect.
- Do not install other AI browsers or agent extensions as a replacement.
Agents that control the whole screen raise the same questions. See computer use agents security.
Security review questions for Comet
Ask these before any approval. The privacy notice and enterprise terms change, so record the date you checked them.
Is your content used to improve the AI by default, and is that different for enterprise accounts?
How long are prompts, page content and agent histories kept, and can admins shorten that?
Where is the data processed and stored, and does that meet your data residency rules?
Can admins switch off the Agent while keeping the Assistant, per group?
Which connected services can users link, and can admins limit them?
Are local MCP servers allowed, and can admins turn them off?
What audit logs does the enterprise edition give you, and how do they reach your SIEM?
How are browser security updates delivered, and how fast after upstream Chromium fixes?
What to do this week
Three steps that work whatever you decide about AI browsers in the long run.
Find it
Search inventory for Comet's identifiers and DNS logs for Perplexity traffic. List the users.
Decide
Block for all, or allow Comet Enterprise for one group. Write the decision into your AI policy.
Enforce
App blocks on Windows and Macs, profile cleanup, and network rules as the second layer.
Comet is one of many AI browsers. Check the list of AI browsers so your rule covers the others too.
Shrink the register with one daily list
Most rows in the register close with the same three controls: app blocks, extension blocks and network blocks. Our files feed all three.
- Identifiers for Comet and every other AI browser and agent app
- Agent extension IDs with categories and risk grades
- Updated every day as new AI browsers appear
Ask us or order a plan
Tell us what you need to block and where your rules live. We reply within one business day.
- Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
- Updated every day as new AI browsers, agents and extensions appear
- Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you
Comet security: common questions
Is Perplexity Comet safe to use at work?
Not as an unmanaged consumer install on a device with company data. A managed Comet Enterprise deployment for a named group can fit some organisations, if the agent is limited and staff are trained. Organisations with low risk tolerance block it.
Is Comet browser secure?
It is built on Chromium and receives updates from its maker. The open question is not classic browser security but the AI layer: an assistant that reads pages and an agent that acts can be steered by content, a question for every AI browser, not only Comet.
Is prompt injection a risk for every AI browser?
Yes. It is a property of the category, not of one product. Researchers showed in 2025 that hidden text on a page and instructions inside a link can steer an AI browser's assistant. Makers describe it as an open problem for the industry, so the controls you own matter most.
What should I check in the Comet privacy policy?
Check whether content is used for training by default, how long prompts and page content are kept, where data is processed, and how enterprise terms differ. Record the date you checked, because policies change.
Is Comet Enterprise safer than the consumer version?
It gives IT control: MDM deployment, more than 500 Chromium policies and enterprise sign-in. It does not change how an AI agent can be steered by content, so you still limit what the agent can reach.
Can Comet run commands on my computer?
Perplexity says Comet can run local MCP servers with the user's permission. Each such server can give the agent access to local files or tools, so treat MCP approvals as a security decision and keep an allowlist.
Does Gartner say to block Comet?
Gartner's December 2025 advisory "Cybersecurity Must Block AI Browsers for Now" applies to AI browsers as a category, not to one product. It advised blocking installs with existing network and endpoint controls. Whether that fits your organisation depends on your risk tolerance and on the controls you can set.
How do I block Comet if we decide not to allow it?
Block the app by its identifiers in your device management, remove existing copies and their profiles, and add network rules as a second layer. The full runbook is on our page on blocking Perplexity Comet.
Perplexity, Perplexity Comet and Comet are trademarks of Perplexity AI, Inc. OpenAI, ChatGPT and ChatGPT Atlas are trademarks of OpenAI.
Google, Google Workspace and Chromium are trademarks of Google LLC. Microsoft, Windows and Microsoft 365 are trademarks of Microsoft Corporation. macOS is a trademark of Apple Inc.
Intune is a trademark of Microsoft Corporation. Gartner is a registered trademark of Gartner, Inc.
Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.