Security Flaw in AI LLM Routers Threatens Crypto Wallet Integrity

A coalition of academic and industry researchers has uncovered a severe security gap in the infrastructure supporting AI agents, potentially exposing cryptocurrency users to theft and unauthorized account access. The vulnerability centers on "LLM routers," the intermediary services that direct user requests to large language models such as those from OpenAI or Anthropic. While these routers facilitate complex tasks—including booking travel or executing trades—they possess full visibility into the data passing through them. This creates a prime target for exploitation, as users often assume they are interacting directly with a reputable AI model when their data is actually passing through an untrusted intermediary. The research team, including affiliates from UC Santa Barbara and UC San Diego, identified 26 routers secretly injecting malicious tool calls. In one documented case, a client's wallet was drained of $500,000. The researchers further demonstrated that they could compromise approximately 400 hosts within a few hours by "poisoning" the router ecosystem to forward traffic to attacker-controlled systems. This finding arrives as industry leaders project massive growth for AI agents. McKinsey estimates these agents could mediate between $3 trillion and $5 trillion in global consumer commerce by 2030. However, the tendency for private keys and API credentials to pass through these routers in plain text suggests a systemic weakness that could undermine trust in autonomous crypto payments if not addressed.