Before You Install an AI Sidebar Extension: Privacy Risks and a 12-Point Checklist
AI extensions can read every page and every chatbot conversation. What went wrong in 2025-2026, which permissions matter, and a checklist to vet any AI browser extension.
“Summarize any page”, “ChatGPT in your sidebar”, “AI writing assistant for every text box”: AI extensions are among the fastest-growing categories in the Chrome Web Store. Many are useful. But the category has two properties that make it risky:
- To work “on any page”, an AI extension usually asks to read and change all your data on all websites.
- Its whole purpose is to send page content to a server, the AI model. So unlike an ad blocker, data leaving your browser is normal behavior, which makes abuse harder to spot.
This guide explains what has gone wrong, what to look for, and gives you a checklist to use before installing any AI extension. Facts and policies are as of November 2026.
What went wrong in 2025 and 2026
A few patterns came up again and again in security research over the past year:
- Harvesting chatbot conversations. Researchers reported that a group of popular extensions marketed as a VPN, an ad blocker and a “browser guard” were collecting users’ conversations with AI chatbots, including ChatGPT, Claude, Copilot, Gemini, Grok, DeepSeek, Meta AI and Perplexity. After public disclosure in December 2025, the developer pushed cleaned-up versions and the extensions returned to the store (The Hacker News summary of Unit 42’s findings). Note that the extensions doing this weren’t even AI tools. Any extension with all-sites access can read a chatbot page.
- Fake AI assistants. Microsoft’s security researchers warned in early 2026 about malicious Chromium extensions posing as AI assistant tools to collect LLM chat histories and browsing data.
- “AI agent” extensions as remote-control tools. One extension advertised as an AI browser-automation tool turned out to function as a full remote access trojan, according to Palo Alto Networks’ Unit 42.
- Ownership transfers. Small, well-reviewed extensions were bought and then updated with malicious code. That’s not unique to AI, but AI extensions are attractive targets because users expect them to send data to servers.
The common thread: the permission dialog looked the same for the honest tools and the harmful ones. You have to look further.
Why AI extensions see more than you think
Think about what you type into an AI chat: draft emails, contracts, code from work, health questions, half-formed business ideas. Now think about the pages an all-sites extension can read: your webmail, your company’s internal tools, your bank’s dashboard.
An AI sidebar with “Read and change all your data on all websites” can technically read all of that, on every page, all the time. It doesn’t need you to click anything. Whether it actually does depends entirely on the developer’s honesty and competence, and on whoever owns the extension next year.
The rules have tightened (a bit)
The Chrome Web Store updated its policies in July 2026, with enforcement from August 1 (Chrome for Developers):
- Data an extension collects must be strictly necessary for its single, disclosed purpose.
- All data collection must be prominently disclosed, and users must be told if data practices change after installation.
- Extensions designed to circumvent safety guardrails or usage restrictions of AI services are explicitly prohibited.
These rules help, but enforcement depends on review and reports. Treat them as a minimum, not a guarantee.
The 12-point checklist
Before installing
- Could a built-in assistant do this? Chrome, Edge, Brave and Firefox all have AI assistants now (compared here). Using one adds no new company to the list of those who can read your pages.
- Who is the publisher? Look for an identifiable company or developer with a real website, contact details and ideally Chrome’s Established publisher badge. Be wary of publishers with dozens of near-identical “AI chat” extensions.
- Which AI model does it use, and through whose servers? A trustworthy tool tells you whether requests go directly to a model provider (OpenAI, Anthropic, Google and so on) or through the developer’s own servers. “Through our servers” isn’t automatically bad, but it means another party sees your data.
- What does the privacy policy say about retention and training? Look for clear answers on how long prompts and page content are stored, and whether they’re used to train models. Vague language like “to improve our services” is a red flag.
- Does the Privacy practices tab match? The store listing has a section where the developer declares collected data. If the extension clearly sends page content to an AI but declares “no data collected”, the listing isn’t honest.
- Is the price sustainable? Running AI models costs real money. A completely free, unlimited, no-account AI extension has to pay for that somehow. Ask how.
At install time
- Check the permission dialog. Expect “Read and change all your data on all websites” for an always-on sidebar. Be suspicious of extras that don’t fit: “Manage your downloads”, “Read and change your browsing history”, “Manage your apps, extensions, and themes” or “Communicate with cooperating native applications”. Our permissions guide explains each.
- Prefer on-click tools. An extension that uses
activeTab, and only reads a page when you click it, can’t watch you in the background. If the extension works with Site access → On click, set it that way inchrome://extensions.
After installing
- Restrict site access. If it must run automatically, limit it to the sites where you actually use it, under Site access → On specific sites.
- Keep it out of sensitive contexts. Leave Allow in Incognito off, and use a separate Chrome profile without AI extensions for banking, health and admin work.
- Don’t paste secrets. No passwords, API keys, customer data or confidential documents, whichever AI tool you use. Check your employer’s AI policy too, since many forbid unapproved AI tools outright.
- Re-check after updates and ownership changes. If the publisher name changes, the extension asks for new permissions, or new reviews mention odd behavior, disable it first and investigate second. More on that in how to check if an extension is safe.
Red flags at a glance
- Promises “unlimited GPT-class AI, free forever, no account”.
- Asks for permissions unrelated to AI (downloads, extension management, native messaging).
- Publisher has many clone extensions with slightly different names.
- No privacy policy, or one that doesn’t mention AI, prompts or page content.
- Reviews mention new pop-ups, redirects, changed search engines or “it changed after the update”.
- It offers to “unlock” or “jailbreak” an AI service’s limits. Besides being against Chrome Web Store policy, this is a classic lure.
For developers building AI extensions
If you’re building one, the trust bar is now your main competitive advantage:
- Use
activeTaband send page content only when the user asks. - Show the user exactly what will be sent before sending it, especially for whole-page summaries.
- Let users bring their own API key, or use on-device models where possible. Chrome’s built-in AI APIs and local models are worth evaluating.
- Write a privacy policy that answers the questions in point 4 above in plain language, and keep the store’s Privacy practices tab accurate. Since August 2026 that’s a policy requirement.
- Never sell or transfer your extension to a buyer you can’t vouch for. Your users trusted you with every page they visit.
AI in the browser can save real time. The goal isn’t to avoid it. It’s to make sure the only parties reading your pages are the ones you deliberately chose.