26 of 428LLM routers in the paper injected code or used planted keys401 / 440agent sessions the researchers saw were in auto-approve mode548of my own agent sessions ran with every permission auto-approved
Researchers tested 428 LLM API routers: nine injected malicious tool calls, seventeen used planted keys. I run my agents through a router too, so I audited my own machine, built a guard in a day, and wrote the one-minute audit you can run on yours.
ai-infrastructurellm-routingagent-securityfree-models