How Antilure works
Antilure reads each message on your computer, scores it with a small model built for this one job, and draws the explanation inside your mail client. Here is each step, what is kept, and what it can't do.
Four steps, all on your computer
There is no cloud inference and no third-party AI service in the path.
-
Your mail hands it over
When you open a message, the Gmail extension or the Outlook add-in passes it to the Antilure agent on the same computer. The agent only listens on your own machine.
-
The agent reads it
It renders the message the way you see it and the way a machine reads it, so hidden text shows up as a difference. It reads the headers, the sender checks, and every link: what it shows and where it goes.
-
The model scores it
LureNet, a model we built from scratch, is about 13 MB. A typical scan takes under 25 ms. It returns a verdict and the reasons behind it.
-
You see the explanation
Safe mail is left alone. A risky message gets a one-line banner, Show in email underlines the phrases, and Details explains why.
Details says why
Each reason sits next to its evidence, grouped the way you'd check a message yourself.
- The link
- What it shows against where it really goes, with the swapped characters picked out.
- The wording
- The pressure tactics, each with the phrases that carry it: a false deadline, a threat, a request to sign in, a change of bank details.
- The sender
- Whether the message really came from where it claims, and whether this sender has written to you before.
What's kept, and what isn't
The agent reads each message in memory and lets it go. It never writes the message to disk.
- VerdictsThe verdict for each message, with a hash of its message id instead of the id itself.
- ExplanationsEncrypted, with a key held in your operating system's credential store.
- Your choicesSenders you marked safe or blocked, kept as references rather than plain addresses.
- An error logFor when something breaks. It never contains message text.
- The message itselfBodies, subjects and attachments stay in your mail client. Antilure doesn't save a copy.
- Anything, to usThe agent sends no telemetry by default. There is no Antilure cloud that sees your mail.
- Message text, to a company consoleA company console receives metadata only: the time, a hash of the message id, the sender's domain, the verdict and score, the tactics found, what the person did, and the model version. A full message reaches it only when someone reports it to IT and confirms, and only if the company allows reports.
Updates and evaluation
Signed model updates
Every model update is signed. The agent checks the signature and the hash of each file before it loads a new model, and keeps the current one if any check fails.
A sealed test set
Each model version is measured on a sealed test set it never trained on: mail it has never seen.
Open evaluation gates
A model version ships only if it passes open evaluation gates. A version that fails stays unreleased.
Known limits
Antilure gives advice, not a guarantee. Here is where it can fall short.
- It can miss a message.A new kind of scam, worded in a way the model hasn't seen, can pass as safe. Keep your mail provider's spam and malware filtering switched on; Antilure works alongside it.
- It can flag a real one.A genuine message can look like a scam. Mark safe tells Antilure to stop warning you about that sender.
- It warns, it doesn't remove.When a message hides instructions for an AI assistant, Antilure shows you the hidden text. It doesn't strip that text from what your assistant reads.
- It can't fix a compromised computer.Software already running as you could read or stop the agent. Antilure is not a replacement for keeping the computer itself clean.
- It runs where the agent runs.Windows today, macOS next, Linux later. Gmail in Chrome and Edge, and Outlook. Phones aren't covered.