A lot of people switch on a VPN, see the lock icon in the menu bar, and assume they are now private everywhere online, right down to the prompts they type into ChatGPT, Claude, or Gemini. Others make the opposite mistake and write VPNs off completely, figuring that because websites already use HTTPS and AI tools collect everything at the account level anyway, encrypting the network link no longer matters.
Neither view holds up once you trace where a request actually goes. A VPN solves a transport problem between your device and the wider internet. AI companies collect data inside the application, after you have logged in and handed the text over yourself. Keeping those two layers separate in your head is the easiest way to avoid expecting one tool to do another tool's job.
What a VPN still fixes on the wire
HTTPS protects the body of a web request, but the network between your laptop and the rest of the internet still leaks plenty of context. Without a VPN, whoever runs the coffee-shop Wi-Fi, the hotel network, the coworking router, or your home broadband can still see the DNS lookups your machine makes, the Server Name Indication (SNI) headers in TLS handshakes, the IP addresses of the servers you talk to, when and how much data you move, and your real IP address and rough location.
A well-run VPN closes that gap. Wrapping your traffic in an encrypted WireGuard tunnel, such as NordVPN using its NordLynx protocol, and routing DNS queries through that same tunnel means the local network only sees a single stream of encrypted traffic going to one server. That shuts down packet snooping and rogue hotspots on shared Wi-Fi, stops your ISP from logging every domain you visit or throttling specific traffic, and keeps sites where you are not signed in from tying your visits together by your home or roaming IP address.
Why the tunnel stops at the AI front door
Follow what happens one step further along the line. Your request leaves the VPN server, crosses the public internet over HTTPS, and arrives at OpenAI, Anthropic, Google, Microsoft, or one of the hundreds of smaller apps built on their APIs. At that point the VPN has done everything it was built to do, and the AI provider still gets everything you sent.
That happens because you are signed in. Almost every AI assistant and coding tool requires an account, usually linked through Google or Microsoft single sign-on, a phone number, and a billing card. Sending your traffic through Zurich or Frankfurt changes the IP address in an access log, while your session token tells the provider's backend who you are anyway.
More importantly, the AI company is the other end of your encrypted connection. Transport encryption is there to stop strangers in the middle from reading your traffic on the way to a server. When you paste a contract, a customer spreadsheet, or a block of source code into a prompt, the AI vendor is that server. It decrypts your request so the model can run, and it stores the conversation under the terms of your account. On most free and consumer plans, those terms let the vendor keep prompts and outputs and fold them into future model training unless you turn training off in settings.
A lot of AI collection also starts before a packet ever hits the network card. Browser sidebars that summarise web pages, meeting bots that transcribe calls, and operating-system recall tools read what is on your screen or in your clipboard locally, then upload structured summaries over your signed-in session. A VPN just carries that upload safely to the vendor's cloud.
The quieter leak: account sprawl across small AI tools
Separating network traffic from account identity points straight to a second problem. Over the last couple of years, most developers, marketers, and founders have not stuck to one or two major AI providers. They have opened accounts on twenty or thirty smaller wrappers: PDF readers, slide generators, audio transcribers, SQL helpers, and browser plugins. Plenty of those tools are thin front ends bolted onto a third-party model API, built in a weekend by tiny teams with very little security oversight.
Signing up for each of them with your main work email and a reused password, or clicking "Continue with Google" without checking what it asks for, scatters your identity across dozens of small databases and puts your real accounts at risk when one of those vendors gets breached. That is an identity problem rather than a network one, and it is what a password manager like NordPass is built for. Giving every external service a unique random password and a masked email address keeps a breach at a small AI app from handing an attacker the login to your cloud console or company inbox.
Layer 1 · Transport
NordVPN for untrusted networks
Encrypts traffic in transit with WireGuard-based NordLynx, prevents DNS and IP leaks on hotel, airport, and shared Wi-Fi, and stops local network snooping before your traffic reaches the destination server.
Layer 2 · Identity
NordPass for account isolation
Generates unique passwords, stores passkeys in an XChaCha20-encrypted vault, and provides email masking so testing new AI tools never exposes your primary credentials or links every signup to one address.
Why this lands on organisations, not just individuals
In our security work at HOUSE603, we see the company version of this mix-up almost every month. A business puts a corporate VPN on every laptop, ticks the remote-access box on a client security questionnaire, and assumes its data is safe. Meanwhile, staff sit on that encrypted corporate VPN pasting customer records, board papers, and private source code straight into personal AI accounts.
Under the GDPR and ISO/IEC 27701, pasting identifiable customer or employee information into an unvetted consumer AI tool is a transfer of personal data to an outside processor (or an independent controller, if the vendor trains models on user inputs) with no Data Processing Agreement and no lawful basis. This is also the exact gap ISO/IEC 42001, the management-system standard for artificial intelligence, asks organisations to close: knowing which AI tools touch your data, setting clear rules on what staff may send to external models, and separating consumer accounts from enterprise or API workspaces where training on your data is turned off by contract.
A short checklist for individuals and teams
- Keep a trusted VPN like NordVPN switched on when using hotel, airport, café, or shared Wi-Fi, and treat it as protection for the connection rather than anonymity inside the apps you log into.
- Open the privacy settings of every AI assistant you use and turn off model training on your chats, or move work tasks onto API and business plans that exclude training by default.
- Use a password manager such as NordPass to generate a fresh password and masked email for every third-party AI tool instead of reusing work logins or binding Google and Microsoft single sign-on everywhere.
- Strip names, email addresses, client identifiers, and credentials out of text and documents before pasting them into any consumer prompt window.
- Remove browser extensions and AI sidebars that ask for permission to read data across all websites, since a VPN cannot protect a page that an extension scrapes straight from your open tab.
- Give staff an approved, contractually covered AI workspace so they do not have to sneak work into personal accounts, and record those controls in your ISO/IEC 27701 and ISO/IEC 42001 documentation.
This article is general information from a security and privacy governance practice, not legal advice. HOUSE603 participates in the Nord Security affiliate programme and may earn a commission when readers purchase through our NordVPN or NordPass links, at no extra cost to the buyer.