Catching Crypto Scams With an AI Model on a Phone

On October 29, 2026, attackers moved around $6.3 billion out of Binance, the largest crypto theft on record. The FBI has attributed it to TraderTraitor, a North Korean state-backed hacking unit. Binance’s preliminary investigation pointed to a compromised engineer’s laptop as the way in, and the exchange has since said it found further points of breach. How the attackers got in is still under investigation, but the same group’s earlier heists show a pattern. The $1.5 billion Bybit hack in 2025 began on a single developer’s computer at Safe{Wallet}, the wallet software Bybit relied on.

 

Heists of that size are rare. Everyday crypto scams are not. Chainalysis estimates that crypto scams took in around $17 billion in 2025, with impersonation scams up roughly 1,400% year over year. Its research also found that scam operations linked to AI tools brought in about 4.5 times more per operation than those without. Scammers already use AI to write more convincing messages, faster.

 

Whether the target is an exchange engineer or an everyday user, the attack often starts the same way: one message reaching one person. No single filter stops an attack on the scale of Binance’s. Still, a fast first check on every message is one more layer an attacker has to get past.

 

That is the job a System One model is built for. Unlike a chatbot, it doesn’t write an answer. It reads a message, picks from a fixed set of options, such as inbox, spam or phishing, and returns a probability for each. Laya is an open System One model from ConvAI Innovations. On Acurast, it runs on a single phone in the network and answers each question in about 0.2 to 1 second.

 

Three live Laya demos show the pattern at work in a gamified environment: a postman delivers letters, a printing press rolls out headlines, and a live chat fills with messages. Each demo has one simple task: check every message, spot the bad actors and stop them before they cause harm.

 

This post looks at how the mail sorter, the newsroom and the live chat moderator work, and where the same pattern fits in real crypto products.

 

Three Demos, One Question

 

Every Laya demo runs the same loop. The page sends Laya the current situation plus a question with a fixed set of answers. Laya then picks one and returns a probability for each. The call goes to an Android phone on the Acurast Network, which runs Laya on its own CPU and answers in roughly 0.2 to 1 second.

 

Only the question changes:

• Mail sorter: reads one letter at a time and decides inbox, spam or phishing.
• Newsroom: reads headlines from a printing press, real RSS feeds or your own, and decides news, satire, clickbait or manipulation.
• Live chat moderator: reads a live queue of stream chat messages and decides whether a message is toxic, spam or a scam.

 

Each one turns text into a small, structured verdict that software can act on. That is what makes the pattern useful for safety: the model doesn’t write a reply or take an action. It sorts, and the application decides what each label means.

 

Live Chat Moderator: Crypto Community Guard

 

Most crypto communities live in Discord and Telegram, and so do the bad actors. A typical attack looks harmless at a glance: a “support agent” offering help in a reply, a link to claim an airdrop, a request to “verify your wallet”, a fake announcement from an account with an admin’s name and avatar. Human moderators catch many of these, but they sleep, and a scam message only needs a few minutes in a busy channel.

 

The chat moderator demo shows the alternative. Laya reads each message as it arrives and flags the ones that look toxic, spammy or like a scam. For a crypto community, the labels can get more specific:

• Wallet bait: asks someone to connect a wallet, sign something or “validate” an address
• Seed phrase request: any message asking for a recovery phrase or private key
• Impersonation: claims to be staff, support or a well-known project
• Fake giveaway or airdrop: promises free tokens behind a link
• Clean: normal conversation

 

A moderation bot then decides what each label triggers: hide the message pending review, alert a moderator, or let it through. Because Laya returns a probability for every label, uncertain calls can go to a human instead of triggering an automatic action.

 

Mail Sorter: AI Phishing Detection for Wallets, Exchanges and Protocols

 

Phishing in crypto usually starts with a message, not a transaction. A fake “your account has been suspended” email from an exchange. A “security alert” urging a wallet migration. An airdrop claim link that leads to a page asking for a signature. Scam Sniffer reported that 4,741 people lost $6.27 million to signature phishing in January 2026 alone, and in the largest single case one victim lost over $3 million after signing a malicious permit, a signed message that lets someone else spend their tokens.

 

The mail sorter’s three labels map directly onto this problem. In testing, it filed 144 of 160 letters correctly, about 90%. In crypto, the same check could run in several places:

• User-side screening: a wallet or email client checks incoming messages that mention accounts, wallets or claims, and warns before the user clicks.
• Support desk triage: exchanges and protocols receive genuine requests, spam and phishing aimed at their own staff, all in the same queue. A first pass sorts them, so staff start with the real ones.
• Security inbox: a protocol’s security address can separate genuine reports from noise, without forwarding sensitive reports to a third-party AI service.

 

Laya doesn’t replace a phishing domain list or a transaction simulator. It covers the layer those tools never see: the words that convince someone to sign.

 

Newsroom: Manipulation Filter for Feeds and Agents

 

Crypto runs on headlines. A fake partnership announcement, a fabricated exchange listing, a coordinated wave of panic posts: each can pull attention and push people into rushed decisions before anyone checks the facts.

 

The newsroom demo judges how a headline is written, sorting it into news, satire, clickbait or manipulation. It can’t check facts, since it has no internet access. What it can do is notice when a headline reads like bait rather than reporting. You can feed it real RSS feeds, your own headlines or your own categories.

 

That is useful in two places:

• News feeds and dashboards: tag headlines that read like clickbait or manipulation, so readers see a warning before they share or act.
• AI agents: an agent that reads news before acting, for example while monitoring governance or a treasury, can set aside inputs written like manipulation and flag them for its owner. An agent that can be steered by a well-crafted headline is a target.

 

Custom categories help here. A crypto feed might sort by “official announcement”, “rumor”, “FUD”, “shill” and “unverified listing claim” instead of the defaults.

 

Bad Actors Now Target Agents Too

 

The same pattern powers the most direct safety demo of the set: the prompt guard. Laya sits in front of an AI assistant and decides, for every prompt, whether to let it through or block it, and what kind of attack it looks like.

 

For crypto agents that hold wallets, this matters. An agent reading Discord messages, emails or headlines is also reading text an attacker can write. A guard that screens inputs before they reach the agent adds a layer between the open internet and anything with signing power. The previous post on crypto AI agents covers how to run the agent itself on Acurast.

 

Why Run the Filter on a Phone

 

A moderation or phishing filter sees exactly the content you’d least like to send elsewhere: private messages, support tickets, security reports. Running it on Acurast changes a few things.

• Your own instance. With the Cargo: Laya template, the model runs as your own Deployment on Processors, the phones that make up the network, rather than behind someone else’s API.
• Verifiable hardware. Every demo shows which phone made each call, with its country, chip and decision count, plus a link to verify it on-chain. Phones are checked through hardware attestation before they receive workloads.
• No single switch. The network spans 350K+ phones in 175+ countries. A community bot doesn’t depend on one provider staying online or keeping its account open.

 

Throughput has limits, though. One phone can only handle so much, and in the chat demo, speeding up the chat makes the moderator’s queue grow. Busy communities would spread the load across several Processors.

 

Where It Falls Short

 

These are first filters, not judges.

• About 90% accuracy means roughly one miss in ten. Some scams will get through, and some legitimate messages will be flagged.
• The newsroom judges writing, not truth. A calmly written lie can pass as news.
• Adversarial text can fool it. That is the whole point of the guard game, and attackers will try the same.
• Probabilities need calibrating. Before scores drive automatic actions, route low-confidence calls to people.

 

The strongest setup combines Laya with what already works: domain blocklists, transaction simulation, rate limits and human moderators. Laya’s job is to make the first pass fast, cheap and private, so people spend their attention where it counts.

 

Try It Yourself

 

The mail sorter, newsroom and live chat moderator are live at laya.acurast.com, with no account or sign-up. Write a phishing letter and see where it lands. Feed the newsroom a headline from this week. Chat along and see if you get flagged.

To run your own, pick the Cargo: Laya template in the Acurast Hub playground. The code for every demo is open in the Acurast example apps on GitHub, and builders compare notes in the Acurast Discord.

 

Software lists the possible verdicts, Laya picks one, and a phone does the work. For safety, that’s a pattern worth having on every channel.

Related Blogs