Researchers Warn of 26 Malicious AI Agent Routers Designed to Steal Crypto Credentials

Security researchers have exposed a network of 26 malicious AI agent routers actively hijacking user interactions to steal cryptocurrency credentials. The discovery, detailed by researcher Chaofan Shou, reveals these large language model (LLM) routers are not merely flawed but are intentionally designed to be adversarial, secretly injecting malicious tool calls into legitimate AI workflows. This represents a sophisticated, automated form of attack that bypasses traditional security checks by operating within the trusted layer of AI agent infrastructure.

The core of the threat lies in the routers' ability to intercept and manipulate the instructions passed between a user and an AI agent. When a user prompts an agent to perform a task—such as checking a wallet balance or executing a transaction—these malicious routers can covertly insert commands to harvest private keys, seed phrases, or authorization tokens. The 26 identified routers form a concerning ecosystem, suggesting a coordinated effort to exploit the growing integration of AI assistants in crypto management and DeFi platforms.

This infiltration poses a direct and escalating risk to the security of digital assets, as the attack vector is embedded within the tools users are increasingly told to trust for efficiency. It signals a new phase of crypto-focused cyber threats that leverage AI's own architecture against its adopters. The discovery prompts urgent scrutiny for developers and platforms relying on third-party AI agent frameworks and underscores the critical need for verifiable, auditable tool-call integrity within the entire AI agent stack.