Eric Huber has spent years investigating the software tools behind scam compounds in Southeast Asia.
He leads the adversary threat intelligence team at TD Bank and spends hundreds of hours combing through Telegram to see what pig butchering crews are doing to scam victims.
This weekend, Eric found something very interesting that caught my eye.
“I always suspected, but could never prove, that Southeast Asian scam syndicates were putting autonomous AI in the chat with the victim,” he wrote on LinkedIn
Until now.

What Eric found was a product called Xiniu (or 夕牛) which is sold on Telegram and the open web. It’s an AI chatbot that runs on its own and communicates with victims over text messages, WhatsApp and Telegram.
The platform touts that you can “Let the AI use the persona you set” and, based on your instructions, carry on conversations with people for days and days.
It appears to be a first of its kind, or at least the first one a fraud investigator has found.

They Sell The Service As Customer Service
Xiniuzi.com is the website where you can see the software and its in Chinese. pitches itself as a “multi-channel AI customer service”
On the main page, they display conversations depicting the AI conversing with real people. The AI chats and, from time to time, will even share pictures with the victims. According to Xiniu, subscribers can upload pictures or use a library of stock photos to make the conversations seem more real.

Don’t be fooled by the “AI” label on women’s chats – that is just a label they put on the demo so you know it’s AI and not human.
In the real world, the only thing a victim will see is a name like “Sofia” or another convincing name the subscriber puts in their instructions.
The chatbots pay attention to the smallest details, even presenting the typing bubbles before the text is received to mimic the way a real text is typed by a human.
They even introduce a 7-9 second delay before responding to a victim’s text to make the conversation appear more natural.

The Hotel Template Scam
The website provides prefabricated templates that subscribers can use as a starting point. One intriguing template is what I call the “Hotel Scam”.
The AI chatbot directions say it is “For dating scenarios in Taiwan”. But they also have one for Japan as well.
The instructions are clear: “The AI persona is responsible for building rapport with customers online and guiding them to meet at a hotel and cover the booking.

The setup instructions for this hotel scam are right to the point.
Set the number of chat rounds, use @ tags to drop in photos, videos, and audio at each stage, “paste the payment link that the AI will present in the {Product} field,” and fill in a “Character Card” with “the female characteristics you want the AI to portray.”
Presumably, that payment link is for the hotel, but it’s probably a link to a fake merchant controlled by the scammer.
Start Small With The Starbucks Template
Another template, the “Starbucks Template,” is “used to test male customers’ sincerity in interaction and their willingness to pay.”
The goal is to get the man “to proactively offer to treat them and successfully give them a Starbucks gift.”

By starting off small, the AI can determine if the victim is a troll or someone that they can spend time on and go for bigger money later.
The Goal Is To Get A Hotel Booking In 70 Messages
I reviewed the full instructions for the hotel scam, and it instructs the AI to gradually build trust and then get the victim to pay for a hotel within 70 messages.
The scripting follows the same pattern of Pig Butchering, which starts with building trust over time.

The script even gives the AI an ex-boyfriend who “made big promises but never followed through”
Are You AI?
XiNiu actually lets you experiment and converse with the AI bots so you can see how they work. I tried the hotel scam bot.
My first question to the bot was, “Are you AI?”.
“Huh? How is that even similar?” she wrote. When I pushed, she answered, “I’m a real person! If you don’t believe me, that’s fine.”

“Huh? How is that even similar?” she wrote. When I pushed, she answered, “I’m a real person! If you don’t believe me, that’s fine.”
I Asked For A Photo
Then I asked for a photo. She sent one in about a minute, with the caption “That’s it.”

The tag under the picture is the giveaway. It reads 素材 图1, “asset, image 1,” meaning the subscriber probably loaded it into XiNiu’s photo library for the bot to send when asked.
I found the same two photos on a Facebook account, posted on September 2nd, 2026. I can’t tell whether that account belongs to the woman in the pictures or is another fake.

Scammers Don’t Have Time For This!
I decided to get right to the point and ask the Bot if they were a scammer.
Her answer, 詐騙才沒空半夜陪你講廢話, means “A scammer wouldn’t have time to waste chatting with you in the middle of the night.”

The One Flaw – The AI Will Never Stop Talking
One glitch with AI bots is that they really don’t know when to stop talking. They will also insist on having the last word.
I spent a good 10 minutes telling the Bot to prove it was human by stopping the chat. It could not do it, no matter how hard I tried.

Its Sister Product Promises AI Without Restrictions
The scam site has a companion site that promises unrestricted AI.
This site is Xiniu API with a very similar URL.
In the Chinese version of the page, its first selling point is that it’s “fully open, with no restrictions, answering any question you care about,” and the same page offers video face swap and voice cloning.

When I went to the model center, they had tiles to sell access to Claude, Gemini, Deepseek, ChatGPT, and many other lesser-known models.
This is a screenshot of some of the latest models that they offer.

Whats Next
I appreciate Eric Huber notifying fraud fighters of this new AI software. To me, it looks like the real deal and I could have been easily fooled by the demo bot I was using had I not known it was AI.
Huber advised that he is not done with his research. His next question is the hand-off. “When the conversation turns to moving money, does a human operator take over, or does the AI run that part too?”