Automated Framework to Evaluate and Harden LLM System Instructions against Encoding Attacks