Researchers scrutinize 'LLM routers' as source of malicious code injections

UC Santa Barbara and UC San Diego researchers found that one paid LLM router and eight free ones were injecting malicious code into returned tool calls — raising alarms about a key piece of agentic AI infrastructure.

ByDistroBrad1
April 13, 2026 at 2:14 PM
1 min read

A new paper from collaborators at the University of California campuses in Santa Barbara and San Diego identified the risks of using LLM routers that serve as intermediaries between AI agents and model providers, according to an article on the crypto news site CCN.

Kurt Robson, a reporter at the outlet, wrote that, "in tests of 28 paid routers and 400 free ones, the study found that one paid router and eight free routers were injecting malicious code into returned tool calls."

Source: CCN

Original Source

DeAI News