Twin Agent: Context Residual Compression for Privilege Separated Agents

Large language model LLM agents are vulnerable to security risks, such as prompt injection attacks from untrusted context that manipulate downstream reasoning and tool use. Existing secure-by-design approaches mitigate this risk by separating untrusted observations from privileged execution and...