AI tool claims 97% efficacy in preventing ‘address poisoning’ attacks

Crypto cybersecurity firm Trugard and onchain trust protocol Webacy have developed an artificial intelligence-based system for detecting crypto wallet address poisoning.

According to a May 21 announcement shared with Cointelegraph, the new tool is part of Webacy’s crypto decisioning tools and “leverages a supervised machine learning model trained on live transaction data in conjunction with onchain analytics, feature engineering and behavioral context.”

The new tool purportedly has a success score of 97%, tested across known attack cases. “Address poisoning is one of the most underreported yet costly scams in crypto, and it preys on the simplest assumption: That what you see is what you get,” said Webacy co-founder Maika Isogawa.

Crypto address poisoning is a scam where attackers send small amounts of cryptocurrency from a wallet address that closely resembles a target’s real address, often with the same starting and ending characters. The goal is to trick the user into accidentally copying and reusing the attacker’s address in future transactions, resulting in lost funds.

The technique exploits how users often rely on partial address matching or clipboard history when sending crypto. A January 2025 study found that over 270 million poisoning attempts occurred on BNB Chain and Ethereum between July 1, 2022, and June 30, 2024. Of those, 6,000 attempts were successful, leading to losses over $83 million.

Related: What are address poisoning attacks in crypto and how to avoid them?

Trugard chief technology officer Jeremiah O’Connor told Cointelegraph that the team brings deep cybersecurity expertise from the Web2 world, which they’ve been “applying to Web3 data since the early days of crypto.” The team is applying its experience with algorithmic feature engineering from traditional systems to Web3. He added:

The newly developed system instead leverages machine learning to create a system that learns and adapts to address poisoning attacks. O’Connor highlighted that what sets their system apart is “its emphasis on context and pattern recognition.” Isogawa explained that “AI can detect patterns often beyond the reach of human analysis.”

Related: Jameson Lopp sounds alarm on Bitcoin address poisoning attacks

O’Connor said Trugard generated synthetic training data for the AI to simulate various attack patterns. Then the model was trained through supervised learning, a type of machine learning where a model is trained on labeled data, including input variables and the correct output.

In such a setup, the goal is for the model to learn the relationship between inputs and outputs to predict the correct output for new, unseen inputs. Common examples include spam detection, image classification and price prediction.

O’Connor said the model is also updated by training it on new data as new strategies emerge. “To top it off, we’ve built a synthetic data generation layer that lets us continuously test the model against simulated poisoning scenarios,” he said. “This has proven incredibly effective in helping the model generalize and stay robust over time.“

Magazine: Crypto-Sec: Phishing scammer goes after Hedera users, address poisoner gets $70K