AI System Developed to Combat Address Poisoning Attacks in Cryptocurrency

Companies specializing in cybersecurity, Trugard and Webacy, have developed an AI system designed to identify «poisoned» crypto addresses, as reported by Cointelegraph.

This tool utilizes a supervised machine learning model, which was configured using real-time transaction data. It also incorporates on-chain analysis, feature engineering, and behavioral context.

In tests conducted on known attack cases, the system achieved an accuracy rate of 97%.

«Address poisoning is one of the most underestimated yet costly frauds in the crypto industry. It exploits a simple assumption: what you see is what you get,” stated Maika Isogawa, co-founder of Webacy.

«Poisoning» is a type of fraud where attackers send small amounts of cryptocurrency to a victim from an address similar to the real one. Typically, the first and last characters of the wallet address match, which is what users focus on most when transferring funds.

The main goal of the attack is to trick the victim into sending money to the attacker. Generally, those who copy the address from their transaction history are more susceptible to this tactic.

Between July 1, 2022, and June 30, 2024, there were over 270 million attempts at «poisoning» in the BNB Chain and Ethereum networks, according to reports. Out of these, 6,000 attempts were successful, enabling fraudsters to earn over $83 million.

Jeremiah O’Connor, the Chief Technology Officer at Trugard, emphasized that the team adapted experiences from Web2 cybersecurity to the realm of Web3, applying proven methodologies to the new environment.

“Most existing attack detection systems in Web3 rely on static rules or basic transaction filtering. These methods often lag behind the evolving tactics, techniques, and procedures used by attackers,” he remarked.

The new system leverages machine learning to learn and adapt to «poisoning» attacks.

“AI can detect patterns that are often beyond human analysis,” Isogawa noted.

O’Connor added that Trugard generated synthetic data for the AI to model various types of attacks, and then utilized supervised learning—training the model with labeled data.

The neural network is continuously improved based on new information as advanced strategies emerge.

“Additionally, we created a synthetic data generation layer that allows us to continuously test the model against simulated poisoning scenarios. This has proven incredibly effective in keeping the neural network reliable over time,” O’Connor reported.

It is worth noting that in March, «poisoning» of crypto addresses yielded hackers $1.2 million in just three weeks. On February 20, one victim lost $763,662 due to address poisoning.