Prompt Injection Cannot Be Cured by Filters: How to Isolate Untrusted Text Using Architectural Patterns

Prompt injection is a structural defect in large language models where untrusted text is executed as an instruction. Filters and system prompts are ineffective: even the best defenses are breached in 90% of cases under adaptive attacks. Working approaches—Dual-LLM, CaMeL, and the Two Meta rule—build isolation rather than trying to distinguish good from evil.

OpenAI Anthropic Google DeepMind Microsoft

Safety and Alignment in the Era of Long-Running Models

OpenAI shares its experience deploying long-running AI models, highlighting new safety risks, observed failures, and improved protective measures due to iterative deployment.

OpenAI