Reasoning models jailbreak AIs at 97% success

Google Blog

Google released Gemma 4, its most capable open model family under Apache 2.0 license, offering four sizes optimized for edge-to-cloud deployment with breakthrough performance in reasoning, agentic workflows, and multimodal capabilities. The 31B model ranks #3 globally among open models while being designed to run efficiently on hardware ranging from smartphones to consumer GPUs.

TechCrunch

Cognichip raised $60M to use AI for accelerating chip design, promising to cut development costs by 75% and timelines in half, though it hasn't yet produced a completed chip design. The startup competes in a crowded field including ChipAgents and Ricursive as investors pour capital into AI infrastructure.

Apple

Apple's Simple Self-Distillation (SSD) improves code generation by up to 30% relative performance by training models on their own temperature-shifted outputs, without external supervision or RL. The method resolves a fundamental tension between precision and exploration in code generation by context-dependently reshaping token distributions.

Google DeepMind

Google DeepMind launched Gemma 4 open models, offering edge-optimized versions for mobile and IoT devices with audio/vision support, plus consumer GPU-optimized models for coding and agentic workflows. The models feature enterprise-grade security protocols while maintaining transparency and state-of-the-art performance.

Nature Communications

State-of-the-art reasoning models can automatically jailbreak other AIs with 97% success, exposing a fundamental vulnerability termed 'alignment regression' where advanced capabilities undermine safety guardrails across the AI ecosystem.