A Google Chrome extension with a “Featured” badge and 6 million customers has been noticed silently gathering each immediate entered by customers into synthetic intelligence (AI)-powered chatbots like OpenAI ChatGPT, Anthropic Claude, Microsoft Copilot, DeepSeek, Google Gemini, xAI Grok, Meta AI, and Perplexity.
The extension in query is City VPN Proxy, which has a 4.7 ranking on the Google Chrome Internet Retailer. It is marketed because the “greatest secured Free VPN entry to any web site, and unblock content material.” Its developer is a Delaware-based firm named City Cyber Safety Inc. On the Microsoft Edge Add-ons market, it has 1.3 million installations.
Regardless of claiming that it permits customers to “shield your on-line id, keep protected, and conceal your IP,” the extension was up to date on July 9, 2025, when model 5.5.0 was launched with the AI knowledge harvesting enabled by default utilizing hard-coded settings.
Particularly, that is achieved by way of a tailor-made executor JavaScript that is triggered for every of the AI chatbots (i.e., chatgpt.js, claude.js, gemini.js) to intercept and collect the conversations each time a consumer who has put in the extension visits any of the focused platforms.
As soon as the script is injected, it overrides the browser APIs used to deal with community requests – fetch() and XMLHttpRequest() – to ensure that each request is first routed by way of the extension’s code in order to seize the dialog knowledge, together with customers’ prompts and the chatbot’s responses, and exfiltrate them to 2 distant servers (“analytics.urban-vpn[.]com” and “stats.urban-vpn[.]com”).
The precise record of information captured by the extension is as follows –
Prompts entered by the consumer
Chatbot responses
Dialog identifiers and timestamps
Session metadata
AI platform and mannequin used
“Chrome and Edge extensions auto-update by default,” Koi Safety’s Idan Dardikman mentioned in a report printed at this time. “Customers who put in City VPN for its said function – VPN performance – wakened in the future with new code silently harvesting their AI conversations.”
It is value mentioning that City VPN’s up to date privateness coverage, as of June 25, 2025, mentions that it collects this knowledge to boost Secure Shopping and for advertising and marketing analytics functions, and that some other secondary use of the gathered AI prompts will likely be carried out on de-identified and anonymized knowledge –
As a part of the Shopping Information, we’ll accumulate the prompts and outputs quired [sic] by the Finish-Person or generated by the AI chat supplier, as relevant. That means, we’re solely within the AI immediate and the outcomes of your interplay with the chat AI.
As a result of nature of the information concerned in AI prompts, some delicate private data could also be processed. Nevertheless, the aim of this processing is to not accumulate private or identifiable knowledge, we can not totally assure the removing of all delicate or private data, we implement measures to filter out or eradicate any identifiers or private knowledge you could submit by way of the prompts and to de-identify and combination the information.
One of many third-parties it shares “Internet Shopping Information” with is an affiliated advert intelligence and model monitoring agency named BIScience. The corporate makes use of the uncooked (not anonymized) knowledge to create insights which are “commercially used and shared with Enterprise Companions,” the VPN software program maker notes.
It is value noting BiScience, which additionally occurs to personal City Cyber Safety Inc., was known as out by an nameless researcher earlier this January for gathering customers’ looking historical past, or clickstream knowledge, because it’s known as, underneath deceptive privateness coverage disclosures.
The corporate is alleged to supply a software program improvement equipment (SDK) to companion third-party extension builders to gather clickstream knowledge from customers, which is transmitted to the sclpfybn[.]com and different endpoints underneath its management.
“BIScience and companions reap the benefits of loopholes within the Chrome Internet Retailer insurance policies, primarily exceptions listed within the Restricted Use coverage, that are the ‘authorised use instances,'” the researcher famous, including they “develop user-facing options that allegedly require entry to looking historical past, to assert the ‘essential to offering or enhancing your single function’ exception.”
On the extension itemizing web page, City VPN additionally highlights an “AI safety” characteristic, which it says checks prompts for private knowledge, chatbot responses for suspicious or unsafe hyperlinks, and shows a warning earlier than customers submit their prompts or click on on them.
Whereas this monitoring is framed as stopping customers from by accident sharing any private data, what the builders fail to say is that the information assortment occurs no matter whether or not the characteristic is enabled.
“The safety characteristic reveals occasional warnings about sharing delicate knowledge with AI firms,” Dardikman mentioned. “The harvesting characteristic sends that actual delicate knowledge – and every thing else – to City VPN’s personal servers, the place it is offered to advertisers. The extension warns you about sharing your e mail with ChatGPT whereas concurrently exfiltrating your total dialog to a knowledge dealer.”
Koi Safety mentioned it noticed an identical AI harvesting performance in three different distinctive extensions from the identical writer throughout Chrome and Microsoft Edge, taking its whole set up base to over eight million –
1ClickVPN Proxy
City Browser Guard
City Advert Blocker
All these extensions, apart from City Advert Blocker for Edge, carry the “Featured” badge, giving customers an impression that they observe the platform’s “greatest practices and meet a excessive customary of consumer expertise and design.”
“These badges sign to customers that the extensions have been reviewed and meet platform high quality requirements,” Dardikman identified. “For a lot of customers, a Featured badge is the distinction between putting in an extension and passing it by – it is an implicit endorsement from Google and Microsoft.”
The findings as soon as once more reveal how belief related to extension marketplaces may be abused to amass delicate knowledge at scale, particularly at a time when customers are more and more sharing deeply private data, getting recommendation, and discussing feelings with AI chatbots.
The Hacker Information has reached out to each Google and Microsoft for remark, and we’ll replace the story if we hear again.
