According to Cointelegraph, crypto cybersecurity firm Trugard and onchain trust protocol Webacy have introduced an artificial intelligence-based system designed to detect crypto wallet address poisoning. Announced on May 21, this tool is part of Webacy’s suite of crypto decision-making tools and utilizes a supervised machine learning model. This model is trained on live transaction data, combined with onchain analytics, feature engineering, and behavioral context to enhance its effectiveness.
The tool reportedly achieves a 97% success rate, having been tested across various known attack scenarios. Webacy co-founder Maika Isogawa highlighted that address poisoning is a significant yet underreported scam in the crypto world. This scam involves attackers sending small amounts of cryptocurrency from a wallet address that closely resembles a target’s real address, often with similar starting and ending characters. The aim is to deceive users into mistakenly copying and using the attacker’s address in future transactions, leading to financial losses. A study conducted in January 2025 revealed that over 270 million poisoning attempts occurred on BNB Chain and Ethereum between July 2022 and June 2024, with 6,000 successful attempts resulting in losses exceeding $83 million.
Trugard's chief technology officer, Jeremiah O’Connor, explained that the team applies its extensive cybersecurity expertise from the Web2 domain to Web3 data. This experience includes algorithmic feature engineering from traditional systems, which they have adapted for Web3. O’Connor noted that most existing Web3 attack detection systems rely on static rules or basic transaction filtering, which often lag behind evolving attacker tactics. The newly developed system, however, employs machine learning to create a dynamic system that learns and adapts to address poisoning attacks. O’Connor emphasized the system's focus on context and pattern recognition, while Isogawa pointed out that AI can detect patterns beyond human analytical capabilities.
The machine learning approach involves generating synthetic training data to simulate various attack patterns. The model is trained through supervised learning, where it learns the relationship between input variables and the correct output. This method is commonly used in applications such as spam detection, image classification, and price prediction. O’Connor mentioned that the model is continuously updated with new data as new strategies emerge. Additionally, a synthetic data generation layer has been developed to test the model against simulated poisoning scenarios, proving effective in helping the model generalize and remain robust over time.
