Working with AI

Comet AI Browser Review: Fast, Focused, and Flawed

I tested Comet in late 2025 and recommended it for creative work alongside Brave for sensitive tasks. Six months later, I dropped Comet entirely. What I found, and why I changed my mind.

8 Oct 2025 · 5 min read · By Sophie

Comet AI Browser Review: Fast, Focused, and Flawed

Update, May 2026: I no longer use Comet.

When I wrote this piece, Comet was the most capable browser-based AI assistant I had tested, and what follows still describes what this kind of tool can do. I stopped using it over ownership. Perplexity has Jeff Bezos as an investor, operates under US jurisdiction and faces publisher lawsuits over copyright, and I was no longer willing to route my browsing through it, however good the product.

I now use Brave Leo for passive page analysis and Claude in Chrome for the occasional agentic task, both inside the Brave browser. The fuller account, with each alternative scored on ethics and capability, is in Ditching Perplexity & Comet: A Guide to Ethical AI Alternatives.

I tested Comet in autumn 2025 and recommended it for creative work, with a separate browser for anything sensitive. Six months later I dropped it.

If Chrome has ever turned your laptop into a space heater during a long editing session, you'll see Comet's appeal. Perplexity's AI browser launched publicly in October 2025, promising speed, focus and an assistant that helped without getting in the way. For designers, writers and anyone running operations across several client projects, it delivered: faster than Chrome, lighter on memory, with an assistant that read the page you were on and could draft, summarise or generate without a tab switch.

Then the security research arrived: prompt injection, cross-tab data leaks and session hijacking through malicious URLs, all documented by Brave Software, LayerX Security and other independent teams.

What Comet got right

Pages loaded quickly, tabs stayed responsive through long sessions and the laptop fan stayed quiet. I tested it while updating a five-page client website; layouts loaded cleanly and CSS edits applied immediately.

The assistant sat in a side panel. I used it to summarise competitor websites while drafting proposals, to write short text blocks such as headlines and email templates, to produce CSS snippets, and to answer research questions without leaving the tab. It could also connect to Gmail, Outlook and Google Calendar to search, draft and reply from the browser, and Perplexity and Squarespace announced a native integration with one-click analytics, SEO checks and asset uploads (the official announcement). Moving across from Chrome took a couple of clicks to import bookmarks, extensions and passwords.

For the work it was good at, Comet was the best browser-based AI I had used.

What the security research found

Between August and October 2025, several security teams disclosed vulnerabilities affecting Comet and other AI browsers.

Brave Software found that attackers could hide instructions in web pages using invisible text, HTML comments or screenshots; ask Comet to summarise or act on the page and it could carry those instructions out without warning. LayerX Security demonstrated CometJacking, in which a single malicious URL hijacked the assistant to pull data from connected services such as Gmail or Google Calendar, across open tabs, even if the user never interacted with the malicious site. Brave's October research showed Comet could be tricked by nearly invisible text inside screenshots, which slips past text-based filters.

Cross-tab exfiltration worried me most. With a client project, an inbox and a Google Doc open together, one compromised tab could instruct the assistant to scrape the others and send the data to an attacker's server, with no stolen password involved.

Perplexity patched some of these issues. Both Brave and Perplexity acknowledged that prompt injection remains unsolved across the whole AI browser category, with new attack patterns appearing faster than fixes.

The two-browser setup I used

For about six months I kept Comet for creative work and research and Brave for anything logged in or sensitive. Website editing in Squarespace, Webflow or WordPress, writing, research on trusted sites such as Behance and Dribbble, CSS snippets, outlines and draft proposals went through Comet. Email, Slack, Google Drive, contracts, invoicing, banking, client databases and every other login stayed in Brave. The split kept Comet's speed while keeping client data out of reach of cross-tab exploits, and I used it daily.

Why I left anyway

The security risks were part of it. What ended the arrangement was the question of who I was handing my browsing patterns to. Perplexity has Jeff Bezos as a major investor and operates under US jurisdiction, so the CLOUD Act applies, and it faces publisher lawsuits over its use of copyrighted content. Over the same year I had been cancelling other US-based tools: ChatGPT went, I moved away from Amazon, Chrome gave way to Brave and Gmail to Proton Mail. Keeping Comet stopped making sense alongside those.

The replacement was simpler than I expected. Brave Leo handles passive page analysis, such as summarising an article or pulling out its main points, inside Brave and under Brave's privacy settings. Claude in Chrome handles the agentic tasks I used to give Comet, from a company with a better record on the questions that pushed me away from Perplexity. Both are good enough for my actual work. The full reasoning, with each alternative scored, is in Ditching Perplexity & Comet: A Guide to Ethical AI Alternatives.

If you still want to use Comet

The product is as I described it. If you use it, the two-browser setup is the workable version: Comet for non-confidential creative and research work, a privacy-focused browser such as Brave for everything logged in. Connect as few accounts to Comet as you can, review its permissions monthly and revoke what you don't use.

Never have banking, client invoices, confidential documents or sensitive email open in Comet while you browse untrusted sites. Review every action the assistant proposes before approving it, and if it suggests something you didn't ask for, such as sending an email, stop and check by hand. When researching unfamiliar or competitor sites, don't ask the assistant to read the page; use another browser, or Comet's logged-out mode.

These safeguards reduce the risk without removing it, and none of them answers the ownership question.

Where AI browsers are going

Both Perplexity and OpenAI have acknowledged that prompt injection is unsolved at the protocol level, not only in their own products. Anyone choosing a tool in this category is weighing how far they trust the company holding their browsing patterns, how well their sensitive work is separated from their AI-assisted browsing, and how comfortable they are with the caveats independent researchers have published. My answers led away from Comet. Yours may not.

Back to the Journal