ScamBuster Turns Phishing Emails Into Intelligence by Pretending to Be the Victim
A French engineer built an AI system that replies to scam emails, plays along long enough to extract bank details and phone numbers, then hands the data to investigators.
Key points
- Laurent Giovannoni, a principal software engineer at Filigran, built ScamBuster as part of his engineering thesis at École Polytechnique in France.
- The system has been running in production since November 2025 and will be shown publicly at Black Hat USA 2026 in Las Vegas in August.
- ScamBuster never sends the first email; it only replies to scam messages that arrive in an inbox.
- A well-chosen persona extracts 5.5 times more useful data from a scammer than a poorly matched one.
- The full source code will be released free under the MIT open-source licence on 5 August 2026.
Most companies that receive phishing emails, fake messages designed to trick people into handing over money or passwords, simply delete them. The scammer loses one target but quickly finds another. Nothing is learned, no one is held to account. As we reported on 8 July, filtering a malicious email stops one message, not the criminal behind it.
Giovannoni wanted to change that equation. His answer is ScamBuster: software that writes back.
How does ScamBuster actually fool a scammer?
ScamBuster creates a fake persona, such as a confused widow or a small business owner who seems easy to deceive. The scammer thinks they've found a real victim. They haven't. Behind the replies is an AI, an automated system that holds a conversation using a large language model, meaning a type of artificial intelligence trained to read and write natural human text.
The system keeps the conversation going just long enough for the scammer to reveal what they need: a bank account number, a phone number, a payment website. Once the scammer asks for money, those details are captured and organised into STIX 2.1 and MISP, standard file types that security teams and law enforcement agencies already use to share threat intelligence.
"What looks like dozens of separate scam emails can trace back to the same few numbers and the same payment domain," Giovannoni told Dark Reading. That cluster of details becomes a starting point for a real investigation.
Should you worry about the system being abused?
One rule is built into the architecture and cannot be switched off: ScamBuster never sends the first message. It only responds to emails that arrive in the inbox. "It never contacts anyone first," Giovannoni told Dark Reading. "This rule is built into the architecture, not just set as a policy."
Does the AI actually get better over time?
It does. ScamBuster tracks which personas work best against which types of scam and adjusts on its own. The gap between a well-chosen persona and a poor match isn't a small tuning detail. "It is the difference between a conversation that gives five indicators and one that gives almost none," Giovannoni told Dark Reading, putting that ratio at 5.5 times.
Giovannoni designed ScamBuster to be cheap to run, using GPT-4o-mini and keeping costs close to zero. Operators can swap in a different model, including free open-source ones, by entering a single configuration token at setup.
The code drops on 5 August 2026. Giovannoni says he's already working on versions that would handle voice-call scams, known as vishing, and text-message scams, known as smishing.
What ordinary people and businesses should know. ScamBuster is aimed at organisations with IT teams, not individual users yet. If you receive an unsolicited email asking for payment or personal details, don't reply and don't click any links. Report it to your national fraud authority, then delete it.
