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

arXiv·medium signal

System instructions in agentic AI applications are vulnerable to encoding-based attacks that bypass safety policies and expose sensitive operational context. This paper presents an automated framework to both evaluate and harden system instructions against these attacks, addressing the gap where LLM system prompts that appear robust to natural-language jailbreaks fail against encoded variants. Directly applicable to production agent deployments.