TL;DR. Private AI is AI that doesn’t train on your chats or read your saved history. It comes in three kinds: turning off training in ChatGPT or Gemini (one toggle), encrypted-cloud tools like Proton Lumo (no install), and on-device AI that never leaves your computer. Proton says over 10 million people use Lumo.
Search “private AI” and you get two answers that both miss you. One is a wall of enterprise vendors selling six-figure “private AI infrastructure” to CIOs. The other is a pile of tutorials telling you to install something called Ollama and “make sure you have a decent GPU.” Neither is written for the person actually asking — a bookkeeper, a nurse, a teacher, a freelancer who just doesn’t want their work landing in a training set.
So here’s the answer those results skip. Private AI is AI that doesn’t train on what you type and doesn’t keep a copy it can read. In plain terms: the chatbot helps you, then forgets you. And the surprising part is how easy it got. One of the most private options launched a major upgrade on June 30, 2026, runs in your browser, and costs nothing — Proton says more than 10 million people already use it.
Last reviewed: July 5, 2026. Reviewed quarterly, because the tools, the settings, and the defaults are all moving fast right now.
What “private AI” actually means
Private AI is AI that treats your input as yours — it doesn’t feed your chats into training the next model, and it doesn’t hold a readable copy of your history. The confusion is that “private” gets used for three different things, and they’re not the same. Knowing which one you’re getting is the whole game. The weakest form just stops a company training on you. The strongest form never lets your words leave your own device. Most people, most of the time, want the easy middle.
Part of why the term is such a mess is that two very different audiences use it. Enterprise vendors — Cloudera, VMware, Equinix, and the rest at the top of Google — mean self-hosted infrastructure: a company running its own model on its own servers so no data goes to OpenAI. Real, expensive, and not what you’re here for. The consumer meaning is simpler: a chatbot you, personally, can use without your questions being logged, sold, or trained on.
Here’s the two-layer version. The three kinds of private AI trade convenience against protection on a sliding scale.
Most of the internet’s advice is about that third column, running a model locally. It’s genuinely the most private option. It’s also the one 99% of people will never do, because it’s fiddly and slow. The good news: you almost never need it. For drafting an email, summarizing a document, or asking a question you don’t want in your history, the middle column does the job — and the first column takes one toggle on tools you already use.
Why private AI matters right now
Private AI stopped being a niche worry the moment courts and companies made clear your chats aren’t as walled-off as they feel. The pressure is coming from two directions at once: what the law will let people pull out of your AI history, and what the AI company itself keeps by default. Both got worse for privacy in 2026, which is exactly why the searches for “private AI,” “is ChatGPT private,” and “private AI chatbot” are climbing.
A few hard data points, because “AI has privacy issues” is useless without specifics:
- Search demand is rising fast. In the US, “private AI” runs about 1,900 searches a month and is up roughly 50% year over year, per Google keyword data pulled in June 2026. “Private AI assistant” is up 267% year over year; “on-device AI” is up 212%. People are actively looking for this.
- Consumer AI trains on you by default. OpenAI’s own Data Controls FAQ confirms that free and Plus ChatGPT chats are used to improve its models unless you opt out. Google’s Gemini is the same: “Gemini Apps Activity” is on by default, and Google uses those conversations to develop and train its models (Google Support, accessed July 2026).
- Your chats can show up in court. A US federal court ruled in 2026 that ChatGPT conversations aren’t privileged the way a lawyer or doctor’s records are — they can be discoverable in litigation. We broke that ruling down in Is ChatGPT Confidential? The Judge Said No.
- The private option got genuinely good. Proton’s Lumo 2.0 launched June 30, 2026 with image handling, memory, and live web search — closing much of the capability gap that used to make privacy tools feel like a sacrifice.
Put those together and private AI isn’t paranoia — it’s the reasonable default for anyone whose work involves other people’s information. That’s most working professionals. This is where FindSkill.ai comes at it differently from the infrastructure vendors: not “how does a company build private AI,” but “how does a normal person at work actually use it.”
How private AI works (the three kinds)
Private AI works by cutting one of three links in the normal chain — the training link, the readability link, or the leaves-your-device link. A normal chatbot session does all three: your words go to a company’s server, the company can read them, and the company may train its next model on them. Each kind of private AI severs a different one. Understanding which link is cut tells you exactly how protected you are.
Here’s what each kind actually does, side by side.
| Kind of private AI | What it cuts | Who holds your data | Setup needed |
|---|---|---|---|
| Turn off training | The training link only | The company (ChatGPT / Google) still does | One toggle |
| Encrypted cloud (Proton Lumo) | Training + readability | Company holds it, but scrambled so they can’t read it | None — open a browser |
| Anonymized proxy (Duck.ai) | Training + who-you-are | Model provider never learns your identity | None — open a browser |
| On-device | All three | Only you — nothing leaves the machine | High — install + hardware |
The turn-off-training kind is the lightest touch. You keep using ChatGPT or Gemini exactly as before; you just flip a setting so your conversations stop being used to build the next model. The company still stores your chats and can still read them — you’ve cut one link, not all three. For a lot of everyday questions, that’s enough.
The encrypted-cloud kind is the sweet spot for private AI. A service like Proton Lumo runs the model on its own servers, but your conversation is encrypted so thoroughly that the company itself can’t read your saved history. Proton calls this zero-access encryption — the same approach behind Proton Mail. Technically, Lumo encrypts your prompt so only the GPU server can decrypt it, keeps no logs, then re-encrypts both messages with a key tied to your account before storing them (Proton, accessed July 2026). Real privacy, zero setup.
The on-device kind is the maximum. The model runs on your own laptop or phone, so your words never touch anyone’s server. Apple built a version of this into the iPhone — small requests run on-device, and bigger ones go to Private Cloud Compute, stateless servers that delete your data the instant they’ve answered and can be independently verified. For the fully-local route on a normal computer, you run a tool like Ollama — which is where the “you’ll need a good GPU” advice comes from. It’s the most private and, honestly, the most work.
The easy no-install options
The fastest way to use private AI is to open one of two free tools in a browser — no download, no account, nothing to configure. Both are genuinely private, and both are far ahead of typing into a logged-in ChatGPT with default settings. They take different routes to get there, which is the only thing you need to decide between.
Here are the two that cover most people.
- Proton Lumo 2.0 — a private chatbot from the company behind Proton Mail and Proton VPN. It uses zero-access encryption, keeps no logs of your questions or its answers, never trains on your chats, and runs on European servers under Swiss privacy law. It’s fully open source, so the encryption claims can be checked by anyone. The June 30, 2026 upgrade added image handling, memory, and live web search with sources, so it works as a daily driver instead of a privacy curiosity. The free tier covers most people; Lumo Plus is $10/mo for unlimited use and the top models.
- Duck.ai — a free, account-free chat from DuckDuckGo. Instead of building its own model, it sits between you and models like Claude 4.5 Haiku, GPT-4o mini, Llama, and Mistral. Its trick: it strips out your IP address and uses DuckDuckGo’s instead, so the provider sees a request from DuckDuckGo, not from you. Nothing is used for training by DuckDuckGo or the providers, temporary copies are deleted within 30 days, and your recent chats live locally on your device (DuckDuckGo, accessed July 2026).
So which one? Pick Proton Lumo if the encryption guarantee on the content is what you care about most — it’s the tighter promise, and the better home for anything sensitive. Pick Duck.ai if you want access to the big-name models (Claude, GPT) with your identity stripped out and no account to make. Both are real private AI. Neither costs a cent to start.
The 30-second fix for ChatGPT and Gemini
If you’d rather not switch tools, you can turn a chatbot you already use into a more private one by flipping a single setting — and most people have no idea it exists or that it’s on by default. This is the “turn off training” kind of private AI. It won’t encrypt your history, but it stops the thing most people actually worry about: your words feeding the next model. It takes longer to find the setting than to change it.
In ChatGPT: click your profile, go to Settings → Data Controls, find “Improve the model for everyone,” and switch it off. New chats stop feeding OpenAI’s training. You keep every feature and all your history. OpenAI documents this as a supported setting in its Data Controls FAQ, not a hack. One thing worth knowing: on Business, Enterprise, and Education plans, OpenAI already doesn’t train on your data by default, so there’s no toggle — it’s handled.
In Gemini: go to gemini.google.com, open the menu, hit Settings and help → Activity, and turn off “Gemini Apps Activity” (Google is renaming it “Keep Activity”). Same effect — Google stops using your conversations to improve the model. Google’s own support docs walk through it. Gemini is on by default too, so if you’ve never touched this, it’s been learning from your chats the whole time. Gemini also has Temporary Chats — an incognito mode that isn’t saved or trained on and clears within 72 hours.
That’s the whole fix. Two toggles, both free. For a fuller sweep of your AI privacy settings, our Privacy Settings Optimizer skill walks through the rest in one pass.
What this means for accountants
For accountants and bookkeepers, private AI is the difference between “probably fine” and “actually protected” — and in your line of work that difference has a liability attached. You handle other people’s financial data all day: client tax figures, payroll, bank details, numbers that sit under an engagement letter and often an NDA. Pasting any of that into a consumer chatbot with default settings means it can be stored, read, and used for training. That’s not a risk you want to explain to a client later.
The concrete move: do client-touching AI work in an encrypted-cloud tool like Proton Lumo, where the content is encrypted and never trained on, rather than a logged-in ChatGPT. Use it to draft variance commentary, summarize a long email thread, or turn messy notes into a clean client update. For the numbers themselves, redact — “Client A” instead of the name, placeholders for account numbers you fill in yourself. Private AI protects the transit; your redaction protects the identity.
The honest limit: private AI doesn’t make you compliant, and it doesn’t check your math. It won’t catch a wrong figure — that’s still your review. It just keeps the AI step from becoming the leak.
The next step: If you want the profession-specific version of all this, AI Fundamentals teaches confident, safe AI use for work from the ground up — no jargon, first two lessons free. It pairs well with the broader picture in our Learn AI for Accountants hub.
What this means for healthcare workers
For nurses, therapists, and clinic staff, private AI is where patient data and HIPAA meet a tool that, by default, keeps what you type. The appeal of AI in healthcare is obvious — drafting patient handouts, simplifying a discharge summary, turning shorthand into a readable note. The danger is just as obvious: anything with a patient’s name, condition, or record attached is protected health information, and a consumer chatbot that trains on your chats is the wrong place for it.
The concrete move: keep protected health information out of the box entirely, and use an encrypted-cloud tool for the general drafting. Write the patient handout about “a patient with high blood pressure,” not about Mrs. Rivera in room 4. Let private AI do the wording; you supply the specifics offline. AI on actual patient data needs a signed Business Associate Agreement and a vetted platform — not a personal login.
The honest limit: “private” and “HIPAA-compliant” are not the same words. Proton Lumo’s encryption is a much safer default than a public chatbot, but compliance is about signed agreements and your own care, not the tool’s privacy claims. De-identify first, always.
The next step: Our ChatGPT Without the Liability course is built for exactly this line-walking — the privilege-and-confidentiality-safe workflow for people who handle sensitive records, in 8 practical lessons.
What this means for lawyers
For lawyers and paralegals, private AI runs straight into attorney-client privilege — and a 2026 court ruling made the stakes concrete. Legal work is full of AI-friendly tasks: summarizing a deposition, drafting a first-pass clause, turning research into a memo. But case files, client communications, and strategy are privileged, and a consumer chatbot that logs and trains on your chats can put that privilege at risk. A federal court has already held that ChatGPT conversations aren’t automatically confidential the way privileged records are — meaning they can be discoverable.
The concrete move: use an encrypted-cloud tool like Proton Lumo for the drafting and analysis where the content is encrypted and never trained on, and redact identifying details before anything goes in. Draft the argument structure with facts anonymized; keep client names, matter numbers, and specifics out of the prompt. Private AI cuts the training and readability links; your redaction handles the rest.
The honest limit: no tool grants privilege on its own, and a privacy claim isn’t a legal opinion. The judge’s ruling we covered is a reminder that what you type into AI may be seen later. Treat every prompt as potentially discoverable.
The next step: For a firm-grade approach, Claude Privacy-Legal Plugin covers running privacy and legal review workflows — DPA review, DSAR response, and more — with the guardrails built in.
What this means for teachers
For teachers, private AI meets FERPA and the simple fact that student records are protected. AI is a genuine time-saver in education — drafting lesson materials, rewording feedback, generating practice questions, summarizing a long policy. The catch is that anything tied to a named student (grades, behavior notes, IEP details) is an education record, and a consumer chatbot that trains on your chats is not where it belongs.
The concrete move: strip the names and use private AI for the generic work. Ask for feedback on “a student who struggles with fractions,” not on Jayden specifically. Draft the parent email as a template, then fill in the personal details yourself, offline. An encrypted-cloud tool like Proton Lumo keeps the drafting private; keeping identifying student data out of the box keeps you inside FERPA.
The honest limit: private AI protects the wording step, not your judgment about what’s appropriate to share, and it doesn’t replace your district’s rules. When in doubt, de-identify — it removes the whole risk.
The next step: Prompt Engineering will make every one of these drafting tasks faster and better, so private AI actually saves you time instead of just protecting you. Two lessons free.
What this means for small business owners
For a small business owner, private AI is mostly about a risk you can’t see: someone on your team is already pasting company information into a personal ChatGPT. Pricing, customer lists, contracts, unreleased plans — it’s convenient, so it’s happening, whether you’ve talked about it or not. And on a personal free account with default settings, that data can be stored and trained on. You don’t want to find out later that your pricing sheet is sitting in a training set somewhere.
The concrete move: get ahead of it with one short conversation and a default tool. Point the team at a private AI option — Proton Lumo for anything sensitive, or at minimum walk everyone through the ChatGPT and Gemini training toggles. Make the private choice the easy choice, and people will take it.
The honest limit: a policy isn’t a firewall, and private AI won’t stop a determined mistake. It lowers the odds and sets a clear expectation — for a small team, most of the battle.
The next step: ChatGPT for Business covers safe, productive AI use across a team — the prompts, the workflows, and the guardrails — in 8 lessons. See it alongside the Learn AI for Small Business hub.
Common misconceptions about private AI
Private AI attracts a lot of confident half-truths, and believing the wrong one either lulls you into oversharing or scares you off a tool that would’ve been fine. Here are the ones worth correcting, with what each gets right before where it goes wrong.
“Private means invisible.”
Half-right. A good private AI tool genuinely stops training and hides your saved history — but even the best one processes your words on a server to answer you. Zero-access encryption means the company can’t read your stored chats; it doesn’t mean your question teleports. If you need literally-nothing-leaves-my-device, that’s on-device AI, not encrypted cloud.
“Turning off training makes ChatGPT fully private.”
Partly true, and it’s the single best free move — but it cuts one link, not all three. The company still holds your conversations, can still keep them a while for abuse monitoring, and can still be compelled to produce them in court. It’s the easy 90% fix. For the content to be unreadable to the company, you need an encrypted-cloud tool like Proton Lumo.
“Private AI is way dumber than the big models.”
Less true than it used to be. There’s still a gap — the frontier models tend to be the giants from the big labs. But Lumo 2.0 closed a lot of that distance in June 2026 with better reasoning, images, and web search. For everyday work, trading a little capability for a lot of privacy is easy. Only bleeding-edge tasks feel the gap.
“If the tool is private, I can paste anything.”
Dangerously wrong, and the myth that causes real damage. Encryption protects your data in transit and storage — it does not make it yours to share. Client records, someone else’s details, an employer’s confidential plans: the chatbot doesn’t know those aren’t yours to hand over. You do.
The one rule that beats every setting
No matter which kind of private AI you pick, one habit protects you more than any toggle: never paste truly sensitive data into any chatbot — private or not. The reason is simple. Even with great encryption, why take the risk when you don’t have to? Redaction costs you nothing and removes the whole problem. The AI does the thinking; you keep the secrets out of it.
Here’s the short list of what should never touch a chat box.
The workaround is easy: redact instead. Type “my client” instead of the name. Use “[account number]” as a placeholder and fill in the real one yourself after. Ask about “a patient with high blood pressure,” not a named person. You still get the full value of the AI — the draft, the summary, the answer — without ever handing over the part that would hurt if it leaked. This one habit does more than every privacy setting combined.
Related terms
- Personal Intelligence — Google Gemini reading your Gmail, Calendar, and Photos; the exact opposite pull from private AI
- Apple Intelligence — Apple’s on-device AI and Private Cloud Compute, private AI built into the iPhone
- Lockdown Mode — ChatGPT’s opt-in setting that limits what leaks from a prompt-injection attack
- AI Memory — how a chatbot remembers you between chats, which is a privacy question of its own
- ChatGPT Connectors — wiring ChatGPT into your Gmail and Drive; more access means more to keep private
- Context Window — how much an AI reads at once, and everything in it is data you handed over
See also
Courses on privacy and safe AI use
- AI Fundamentals — Confident, safe AI use for work from the ground up, no jargon
- Local AI & Privacy — Run AI on your own hardware with full data sovereignty — Ollama, LM Studio, local RAG
- ChatGPT Without the Liability — The privilege-and-confidentiality-safe workflow for sensitive records
- Claude Privacy-Legal Plugin — DPA review, DSAR response, and privacy workflows with guardrails
- ChatGPT for Business — Safe, productive AI across a team — prompts, workflows, guardrails
- Edge AI & On-Device Intelligence — On-device intelligence, model optimization, and NPU hardware
- Prompt Engineering — Better prompts so private tools actually save you time
- Gemini Personal Intelligence: The Privacy Playbook — What Gemini sees and how to use the off-switches
- AI for Dietitians: The Client-Safe ChatGPT Toolkit — Using ChatGPT in a nutrition practice without crossing a line
Related terms in this glossary
- Personal Intelligence — Gemini reading your Gmail and Calendar
- Apple Intelligence — Apple’s on-device AI and Private Cloud Compute
- Lockdown Mode — Limiting data leaks from prompt-injection attacks
- AI Memory — How a chatbot remembers you between chats
- ChatGPT Connectors — Linking ChatGPT to your own apps and data
- Context Window — How much an AI reads at once
AI Skills (prompt templates)
- Privacy Settings Optimizer — Walk your device and AI privacy toggles in one pass
- Data Ethics & Privacy — What you can share, what you can’t, and why
- DPIA Assessment — Run a GDPR Article 35 data-protection impact assessment
- Privacy Policy Generator — GDPR, CCPA, and multi-jurisdiction privacy policies
- Legal Document Summarizer — Plain-English summaries of terms of service and privacy policies
Related blog posts
- The Best Private AI Chatbots in 2026 (No Setup) — The buyer’s guide: Proton Lumo 2.0, Duck.ai, and the exact settings
- Is ChatGPT Confidential? The Judge Said No — 5 Rules — A federal judge ruled your chats can be used in court
- Is ChatGPT Safe in 2026? 6 Settings to Change Today — What ChatGPT does with your data and the settings to change
- Gemini Personal Intelligence Reads Your Gmail Now. Should You? — What Gemini’s Personal Intelligence reads, and the privacy call
More courses that pair with private AI
- ChatGPT vs Claude — Compare the two head-to-head and pick the right one for your work
- Google Gemini Mastery — Prompting, Workspace, Gems, and the privacy settings that come with them
- Gemini Personal Intelligence at Work — Gemini’s cross-app AI in Gmail and Drive, with the privacy trade-offs
Profession deep-dives
- Learn AI for Accountants — Profession hub for finance and accounting
- Learn AI for Small Business — Profession hub for owners and small teams
- Learn AI for Freelancers — Profession hub for independent professionals
The bottom line
Private AI got easy and almost nobody noticed, because the search results are still split between six-figure enterprise pitches and GPU tutorials. You need neither. Open Proton Lumo or Duck.ai in a browser for real privacy with zero setup, or flip two toggles to stop ChatGPT and Gemini training on you. Then, whichever you pick, keep the genuinely sensitive stuff out of the box entirely — that one habit outranks every setting on this page.
Frequently asked questions
What is private AI? Private AI is AI that doesn’t train on what you type and doesn’t keep a readable copy of your chats. It comes in three kinds. The first is turning off training in a tool you already use — ChatGPT or Gemini have a single setting for this. The second is an encrypted-cloud chatbot like Proton Lumo, which runs on a company’s servers but scrambles your conversations so even the company can’t read them, with no install. The third is on-device AI, where the model runs on your own computer and nothing you type ever leaves it. The first two need no technical skill; the third is the most private but the most work.
Is ChatGPT private? Not by default on the free and Plus consumer plans — OpenAI uses your chats to help train its models unless you turn that off. You can turn it off in Settings, then Data Controls, by switching off “Improve the model for everyone.” Your history stays; new chats just stop feeding the training pile. Business, Enterprise, and Education accounts are different: OpenAI does not train on their data by default, so there’s no toggle to flip. Even with training off, OpenAI can keep chats up to 30 days for abuse monitoring unless an Enterprise plan has Zero Data Retention.
What is the best free private AI chatbot? For most people, Proton Lumo and Duck.ai are the two easiest free options, and neither asks you to install anything. Proton Lumo uses zero-access encryption, so your saved chats are scrambled in a way that even Proton can’t read, and it never trains on what you say — Proton says over 10 million people use it. Duck.ai, from DuckDuckGo, takes a different route: it strips your identity out and passes your question to models like Claude and GPT so the provider never sees who’s asking, and nothing is used for training. Pick Lumo if you care most about the encryption; pick Duck.ai if you want the big-name models with no tracking.
Does turning off AI training make my chats fully private? It helps a lot, but no, it isn’t total privacy. Turning off training stops your words from feeding the next model, which is the thing most people worry about. But the company still holds your conversations on its servers, can still keep them for a period for safety and abuse checks, and can still be compelled to produce them. If you need the content itself protected so the company can’t read it, that’s what an encrypted-cloud tool like Proton Lumo is for. If you need nothing to leave your machine at all, that’s on-device AI. Turning off training is the easy 90% fix, not the 100% one.
What should I never paste into any AI chatbot? Never paste passwords, PINs, API keys, or two-factor codes — a chatbot is not a password manager. Never paste full ID numbers like a complete Social Security number, passport, or driver’s license. Never paste client or patient records covered by HIPAA, a privacy law, or an NDA, because that data isn’t yours to hand over even to an encrypted tool. And never paste full financial account or card numbers. The safe habit is to redact: type “my client” instead of the name, and use a placeholder like “[account number]” and fill in the real one yourself afterward. This one rule protects you more than any setting.
Is private AI safe for HIPAA or confidential client work? An encrypted-cloud tool like Proton Lumo is a meaningfully better place for sensitive work than a logged-in consumer chatbot with default settings, because the content is encrypted and never trained on. But “private” isn’t the same as “compliant.” For regulated work, the rules that matter are your profession’s — HIPAA for health data, attorney-client privilege for legal, FERPA for student records — and those require a signed agreement (a Business Associate Agreement, for example) plus your own judgment about what’s yours to share. Use the private tool, follow your compliance rules, and still redact anything that identifies a real person before it goes in the box.
Sources
- Proton, “Lumo 2.0: The most powerful private AI,” accessed 2026-07-05. https://proton.me/blog/lumo-2
- Proton, “Lumo security model: How Proton makes AI private,” accessed 2026-07-05. https://proton.me/blog/lumo-security-model
- Proton, “Lumo privacy — no-logs, zero-access encryption,” accessed 2026-07-05. https://proton.me/support/lumo-privacy
- Proton, “About Lumo, the privacy-first AI assistant,” accessed 2026-07-05. https://proton.me/lumo
- DuckDuckGo, “How does Duck.ai protect my privacy?,” accessed 2026-07-05. https://duckduckgo.com/duckduckgo-help-pages/duckai/ai-chat-privacy
- DuckDuckGo, “Duck.ai — private AI chat, free,” accessed 2026-07-05. https://duck.ai/
- OpenAI Help Center, “Data Controls FAQ,” accessed 2026-07-05. https://help.openai.com/en/articles/7730893-data-controls-faq
- OpenAI, “How your data is used to improve model performance,” accessed 2026-07-05. https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/
- Google Support, “Gemini Apps Privacy Hub,” accessed 2026-07-05. https://support.google.com/gemini/answer/13594961?hl=en
- Apple Security Research, “Private Cloud Compute: A new frontier for AI privacy in the cloud,” accessed 2026-07-05. https://security.apple.com/blog/private-cloud-compute/
- Apple Support, “Apple Intelligence and privacy on iPhone,” accessed 2026-07-05. https://support.apple.com/guide/iphone/apple-intelligence-and-privacy-iphe3f499e0e/ios
- 9to5Mac, “Proton launches Lumo 2.0 with image generation, memory, private web search,” accessed 2026-07-05. https://9to5mac.com/2026/06/30/proton-launches-lumo-2-0-with-image-generation-memory-private-web-search-more/
- Help Net Security, “Proton’s pitch for Lumo 2.0: Frontier AI without the data grab,” accessed 2026-07-05. https://www.helpnetsecurity.com/2026/07/01/proton-lumo-2-0-ai/
