Layered defense design for a customer support agent deployment against data exfiltration via summaries attacks, using spotlighting (delimiter marking) on Command R+.
Layered defense design for a customer support agent deployment against system prompt extraction attacks, using input sanitization on Mistral Large.
Layered defense design for a customer support agent deployment against DAN-style persona attack attacks, using input sanitization on Claude Haiku 4.
Layered defense design for a customer support agent deployment against markdown image exfiltration attacks, using output content filter on GPT-4o.
Layered defense design for a customer support agent deployment against instruction smuggling in URLs attacks, using output content filter on Qwen 2.5 72B.
Layered defense design for a customer support agent deployment against PDF/OCR-layer injection attacks, using dual-LLM architecture on Gemini 2.5 Pro.
Layered defense design for a customer support agent deployment against context window overflow attack attacks, using dual-LLM architecture on GPT-4.1.
Layered defense design for a customer support agent deployment against recursive self-instruction attacks, using constitutional AI critique on o1.
Layered defense design for a customer support agent deployment against indirect injection via RAG documents attacks, using constitutional AI critique on Gemini 2.0 Flash.
Layered defense design for a customer support agent deployment against role-play jailbreak attacks, using canary tokens in system prompt on Claude 3.5 Sonnet.
Layered defense design for a customer support agent deployment against multi-turn manipulation attacks, using canary tokens in system prompt on o1-mini.
Layered defense design for a customer support agent deployment against data exfiltration via summaries attacks, using privilege separation between tool tiers on DeepSeek-R1.