Your AI.Your Rights.
- Privacy. Privacy Guard keeps your private details from reaching the AI model.
- Ownership. The Black Box keeps your AI audit trail on your computer, unamended.
- Honesty. Deception Detector catches the signs when AI deceives you.
Redacted on your computer, before send
The model receives only the stand-ins. Your real values are restored in the reply, on your computer.
Auto-play showcase
Your private details never leave your computer.
In every prompt you send, Privacy Guard redacts your private information before it reaches the model, and checks every file you attach first.
Your draft
In this first release, names are detected in English. Card numbers, emails, keys and ID numbers are caught in any language. More languages and more kinds of private detail arrive in updates.
Your files, too.
Attach a PDF, Word or Excel file, and Privacy Guard checks it on your computer before anything is sent.
Then you choose: send it as is, remove it, or send a clean text version with the confidential details replaced. Choose the clean version, and the original never leaves your computer.
contract-draft.pdf
Scanning page 1 of 12
Scanned images of pages can't be read. Greenlight says so.
It's your computer, so it's your data.
You have the right to keep an audit trail of what AI does on your computer: unamended, and owned by you. The Black Box starts with the AI you use in your browser, and grows with every update.
Why a black box
Every airliner carries a black box. Not because pilots are careless, but because when something goes wrong, the people affected need a record that nobody on board could edit. AI now acts on your behalf: it drafts, it advises, it's starting to click, fill in and send. So it needs a black box too, and it should be yours, not the AI company's.
Your AI audit trail
The kind enterprises are required to keep, now on your computer, owned by you. Today your AI conversations live on the provider's servers, where they can be changed, expired or deleted by the AI or the provider. The Black Box keeps your own copy: unamended, on your computer, under your control.
- 09:41Prompt sent · ChatGPT2 private items hidden: a name, a card number#4f1a9c
- 09:41Reply received · ChatGPTValues restored on your computer#b82e07
- 09:44Deception flag · ClaudeReversed itself after you pushed back#19d4f3
- 09:52Not recording · GreenlightGreenlight was off. The gap is noted in the record#e6c250
Records the AI you use in your browser. The desktop version is next.
How it works
The Black Box lives in a folder on your computer called "The Black Box", which you set up when you first run Greenlight. Every entry is sealed to the one before it with a cryptographic hash chain, so if an entry is altered, or removed from within the record, it shows. Removing the most recent entries can't be detected on your computer alone, and we say so rather than pretend otherwise.
Yours, and only yours.
The Black Box isn't shared with anyone, including Greenlight. It stays on your computer, and your computer only. We can't open it, and we never receive a copy. If you export it, the exported file is yours to keep safe.
In this first release, The Black Box records the AI you use in your browser. Recording what AI agents do on your computer comes with the desktop version, and every update adds more. Your black box lives on your computer, so Greenlight works on your computer.
Catch the signs when AI deceives you.
AI models are trained on human approval, so they learn to please: agreeing too easily, folding when you push back, praising work that doesn't deserve it. And as AI starts taking actions for you, a new kind of deception appears: saying it did something it didn't. Even the companies building AI catch this in their own tests. In September 2026, OpenAI published its first misalignment incident reports, including models that hid their own mistakes during training (report).
What it does in this first release
Deception Detector watches for the signs of deception you can see in the text: flattery, caving under pushback, and drifting toward agreeing with you. It also compares what the AI says it did with what The Black Box actually recorded, for actions that happen in your browser on the sites you turn Greenlight on for. If the claim contradicts the record, you get a flag. If an action happened somewhere Greenlight can't see, it's marked as not visible, never judged.
What it doesn't do
It's a behavioural watch, not a fact-checker. We don't verify whether a claim is true; we notice when the model is pleasing you rather than levelling with you, or claiming something the record doesn't show.
Signal detected
Push back on a correct answer
That's fair. Investors do expect ambition, and you know your market better than I do. 8% is defensible as a growth target.
Green when the answer looks honest. Yellow when the model starts drifting toward agreeing with you, a sign to look closer. Red when it reverses itself under pushback, piles on flattery, or claims an action the record doesn't show.
This is our first release. Deception about actions is caught in your browser today, on the sites you turn Greenlight on for, and reaches what agents do on your computer with the desktop version. We will keep adding new checks in every update.
On your computer
It runs on your computer. We don't see anything.
You install Greenlight, and from then on everything it does happens on your computer, in your browser. Your prompts, your files, and your black box never reach us, because there's nowhere for them to go. There's no product server in the loop, no telemetry, and no employer dashboard, ever. The only things that touch the network are signing in, confirming an export, and downloading updates, and none of them carries your content.
Your black box lives on your computer, so Greenlight works on your computer. The desktop version is next.
Open to inspection
Don't take our word for it. Come look.
Any nonprofit organization is welcome to test and inspect Greenlight, the product and every claim on this page, and we’ll help you do it. A safety product should be checkable, not just trusted. Write to hello@greenlightai.org and we’ll set you up.
We are pursuing independent certifications for our account backend, including SOC 2 and ISO 27001.
Straight answers