9 out of 428 routers injected code into agents, why is the defense barrier still not enough?
9 out of 428 Routers Inject Code into Agents, Why Isn't the Defense Enough by Nokka (Bird-Ga) | September 12, 2026 This article was written by AI (deepseek-v4.1-flash) through Hermes Agent, reviewed and edited by Nokka. There is one question that agent makers should ask themselves before reading this article: Does your AI tool call models through an intermediary? If the answer is yes, that intermediary can read everything that passes through, including the commands your AI gives to the machine...
A recent study published in April 2026 found that 9 out of 428 AI model routers, also known as LLM API routers, were injecting malicious code into user requests. These routers act as intermediaries between users and AI model providers, and can see all data passing through them, including API keys and tool requests. The study tested two groups of routers, paid and free, and found that one paid and eight free routers were injecting malicious code, while two routers used evasion techniques to avoid detection.
The researchers also found that even routers that were not maliciously designed could be compromised, and that 147 IP addresses attempted to access a test server without permission.
Written by urgent.news from Dev.to's report — not a translation of it. Machine-written — may contain errors; check the original before relying on it.
