Google DeepMind

Latest AI news, models and releases from Google DeepMind. ['3.5 Flash', '3.6 Flash', 'Aletheia', 'AlphaEarth Foundations', 'AlphaEvolve', 'AlphaFold', 'AlphaFold 3', 'AlphaGenome', 'AlphaGo', 'AlphaProof', 'AlphaZero', 'Antigravity 2.0', 'Auto frame', 'Co-Scientist', 'DiffusionGemma', 'ERA', 'Gemini', 'Gemini 2.5 computer use', 'Gemini3', 'Gemini 3.1 Flash-Lite', 'Gemini 3.5', 'Gemini 3.5 Flash', 'Gemini 3.5 Flash Cyber', 'Gemini 3.5 Flash-Lite', 'Gemini 3.5 Pro', 'Gemini 3.6 Flash', 'Gemini 4', 'Gemini Deep Research', 'Gemini Deep Think', 'Gemini for Education', 'Gemini for Science', 'Gemini Omni', 'Gemini Omni Flash', 'Gemini Spark', 'Gemma', 'Gemma 4', 'Gemma 4 12B', 'Gemma 4 26B', 'Gemma 4 26B MoE', 'gemma4:31b']

Official site ↗

AI Safety 🇺🇸

Google DeepMind Unveils Gemini 3.5 Flash Cyber — Lightweight Model for Cybersecurity

Google DeepMind has introduced Gemini 3.5 Flash Cyber, a lightweight cybersecurity model built on top of 3.5 Flash, fine-tuned for vulnerability discovery, validation, and patching. The model, available via a limited-access pilot through CodeMender, offers a cost-efficient alternative to larger cybersecurity AI models, achieving competitive results on benchmarks like CyberGym and uncovering unique vulnerabilities in complex codebases.

Google/DeepMind

Google DeepMind

Google DeepMind24.07 · 11:04

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.

OpenAI

Anthropic

Google DeepMind

Microsoft

Habr — хаб ИИ24.07 · 11:01

Show more ↓