Researchers claim breakthrough in fight against AI’s frustrating security hole
Here’s how it works. First, the system splits responsibilities between two language models: A “privileged LLM” (P-LLM) generates code that defines the steps to take—like calling a function to get the last email or sending…