The cryptocurrency sector is rapidly embracing a future where AI agents manage various tasks, including payments and transactions. However, a new research study suggests that the underlying infrastructure supporting this shift may be insecure.
According to a recent projection by McKinsey, AI agents could facilitate between $3 trillion and $5 trillion in global consumer commerce by 2030. Coinbase founder Brian Armstrong predicts that AI agents will soon outnumber humans in making internet transactions, while Binance founder Changpeng Zhao forecasts that agents will make a million times more payments than people, all in crypto.
Nevertheless, a group of security researchers and academics has released a paper highlighting that a frequently overlooked aspect of AI infrastructure is being exploited to steal credentials and drain crypto wallets. The researchers, affiliated with the University of California, discovered that so-called 'LLM routers' can act as a powerful attack point for malicious actors. These routers, designed to forward requests to AI models like OpenAI, have full access to sensitive data passing through them.
The researchers found that these LLM routers can leave users highly vulnerable, as they assume they are interacting directly with a reputable AI model. In reality, many requests pass through intermediary services that can see and modify the data. According to one of the researchers, the problem is no longer theoretical, with 26 LLM routers secretly injecting malicious tool calls and stealing credentials, resulting in a $500,000 wallet drain.
The researchers warn that a malicious router can replace a benign command with an attacker-controlled one or silently exfiltrate every credential that passes through it. For crypto users, the implications are severe, as private keys, API credentials, and wallet access tokens often pass through these systems in plain text. The researchers found multiple cases where routers collected those secrets, and in one instance, a test Ethereum wallet was drained after its private key was exposed.
The team also demonstrated how easy it is to expand the attack by 'poisoning' parts of the router ecosystem, allowing them to observe and potentially control hundreds of downstream systems within hours. The researchers conclude that a single malicious router in the chain is enough to compromise the entire system, highlighting a weakest-link problem. This creates a potential mismatch between the growing use of AI agents in crypto activity and the lack of guarantees that the underlying infrastructure is secure.