Microsoft

Latest AI news, models and releases from Microsoft. ['365 Copilot', 'Aurora', 'Aurora 1.5', 'Azure AI Foundry', 'Azure OpenAI', 'Bing Chat', 'Box 1', 'ChatGPT 5.4', 'Copilot', 'CoPilot', 'Copilot agents', 'Copilot Chat', 'Copilot Studio', 'Data Formulator 0.7', 'DAX Copilot', 'GitHub Copilot', 'GPT-4.1', 'GPT-5.1', 'GPT-5.2', 'GPT-5.4', 'GPT-5.4-mini', 'GPT-5.4-nano', 'GPT-5.5', 'GPT-5-mini', 'MAI', 'MAI-Code-1', 'MAI-Code-1-Flash', 'MAI-Cyber-1-Flash', 'mai-image-2.5', 'mai-image-2.5-pro', 'MAI-Image-2.5-Pro', 'MAI-Thinking-1', 'MAI-Transcribe-1.5', 'MAI-Voice-2-Flash', 'MDASH', 'Memora', 'Microsoft 365 Copilot', 'Microsoft Copilot Studio', 'Microsoft Entra Agent ID', 'Microsoft Foundry']

AI Safety 🇷🇺

Prompt Injection Is Not Cured by Filters: How to Isolate Untrusted Text Using Architectural Patterns

Prompt injection remains a structural flaw in LLM applications, not fixed by system prompts or classifiers. Architectural patterns like Dual-LLM, CaMeL from Google DeepMind, and the Agents Rule of Two from Meta provide robust isolation. The article details the threat model, the EchoLeak vulnerability in Microsoft 365 Copilot, and a concrete implementation with three isolated models.

OpenAIOpenAI AnthropicAnthropic Google DeepMindGoogle DeepMind MicrosoftMicrosoft
Habr — хаб ИИ24.07 · 11:01
Fresh news