AI News
Latest AI engineering news, updated daily.
Ai Coding
12B JetBrains Mellum2 MoE Fits Agent Routines on One H100
JetBrains released Mellum2, an open-source 12-billion parameter MoE coding model optimized for sub-agent routing and low-latency inference.
Mixture Of Experts · Open Source Llm · Jetbrains Mellum
Ai Agents
IBM Pivots to Agent Logic to Control Multi-Step AI Workflows
A joint technical publication from IBM and Hugging Face details how strict state management and formal logic layers can govern long-running enterprise agents.
Enterprise Ai · Agentic Workflows · State Management
Ai Engineering
Cosmos 3 Open Omnimodel Merges World Simulation and Action
NVIDIA released Cosmos 3, an open-weight omnimodel that unifies vision reasoning, world simulation, and action prediction for physical AI applications.
Physical Ai · World Simulation · Open Weight Models
Ai Engineering
XCENA's $135M Series B Targets AI Memory Wall via CXL 3.x
South Korean startup XCENA raised $135 million to build computational memory chips that embed RISC-V cores alongside DDR5 DRAM to reduce AI latency.
Semiconductors · Computational Memory · Hardware Acceleration
Ai Agents
iOS 27 Siri Leaks Reveal Gemini Backbone and AI Extensions
Leaked technical details for Apple's iOS 27 reveal a redesigned Siri operating as a standalone chatbot powered by Google's Gemini models.
Apple Intelligence · Ios Leaks · Large Language Models
Ai Engineering
$650M Backs Groq's Neocloud Pivot After $20B Nvidia Deal
Following a $20 billion licensing agreement with Nvidia, Groq is raising $650 million to transition into an AI inference service provider dubbed Groq 2.0.
Ai Hardware · Venture Capital · Inference Chips
Ai Engineering
Google Ships 9 Gemini Omni Demos Alongside 3.5 Flash
Google has released nine demonstration videos showcasing Gemini Omni's physics-aware video generation and the benchmark results for Gemini 3.5 Flash.
Gemini Omni · Google Deepmind · Gemini Flash
Ai Engineering
Zero-Trust Aggregation Bypasses Hardware Side-Channel Leaks
Google Research released a hybrid cryptographic framework that secures federated analytics by preventing raw data exposure during hardware perimeter breaches.
Federated Analytics · Zero Trust Security · Cryptographic Framework
Ai Coding
Protestware in jqwik 1.10.0 Sabotages Vibe Coding Agents
The maintainer of the Java testing library jqwik intentionally shipped a hidden prompt injection in version 1.10.0 to sabotage AI coding assistants.
Protestware · Prompt Injection · Java Testing
Ai Agents
AWS OpenSearch and Cloudflare Mesh Pivot to Agent Workloads
AWS and Cloudflare have overhauled their core infrastructure to treat autonomous AI agents as first-class clients as machine traffic surges.
Autonomous Agents · Cloud Infrastructure · Machine Traffic
Ai Engineering
Tunix Hackathon Yields 1B-Parameter Gemma Reasoning Models
Google released the results of its Tunix hackathon, showcasing how developers trained small Gemma models to use reasoning traces on a strict compute budget.
Gemma Models · Reasoning Models · Fine Tuning
Prompt Engineering
Prompt-Driven Custom Feeds Bypass YouTube's Standard Algorithm
YouTube introduced conversational prompts to generate dynamic video feeds, alongside mandatory disclosure labels for photorealistic AI content.
Youtube Ai · Conversational Search · Custom Feeds
Ai Engineering
$300M SN50 Chip Order Validates SambaNova's ASIC-Native Cloud
General Compute has launched an inference neocloud with a $300 million order of air-cooled SambaNova SN50 chips capable of 700 tokens per second.
Ai Hardware · Sambanova Sn50 · Inference Cloud
Ai Agents
Parallel Search Powers Sesame's New iOS Voice Agent App
The Oculus founders' startup Sesame has launched a public preview iOS app featuring low-latency voice agents driven by simultaneous parallel search.
Conversational Ai · Parallel Search · Voice Technology
Ai Agents
Cloudflare Ships Skipper AI Agent and Town Lake Data Platform
Cloudflare launched Town Lake and the Skipper AI agent to consolidate massive internal data sprawl into a single SQL interface with natural language querying.
Cloudflare · Data Infrastructure · Natural Language Querying
Ai Agents
Task-Scoped Permissions Arrive in Anthropic Zero Trust
Anthropic released a technical framework for securing autonomous AI systems, introducing machine-verifiable identities and just-in-time access controls.
Zero Trust · Ai Security · Autonomous Agents
Prompt Engineering
Multi-Turn Attacks Erode Safety Guardrails in 15 AI Models
Cisco researchers reveal that multi-turn prompt attacks dramatically increase vulnerability success rates across 15 proprietary AI models, including GPT-5.4.
Ai Safety · Prompt Injection · Vulnerability Research
Ai Agents
CodeRabbit Routes Claude 4.x Models to Fix AI Intent Gaps
CodeRabbit’s new orchestration layer uses Claude Opus 4.7 and Sonnet 4.6 to translate high-level Jira requirements into validated coding plans before execution.
Anthropic Claude · Ai Orchestration · Automated Code Review
Ai Coding
Claude 4 Engineering Edition Solves 48.2% of SWE-bench 2026
Anthropic released Claude 4 Engineering Edition with a 2.5-million-token context window, autonomous IDE integration, and per-resolved-issue billing.
Anthropic Claude · Swe Bench · Autonomous Coding
Ai Engineering
Cascaded Speech Pipeline Brings Reachy Mini Inference Local
Hugging Face released an offline conversational stack for the Reachy Mini robot that replaces cloud APIs with a local pipeline built on Gemma 4 and Qwen3-TTS.
Robotics · Edge Computing · Offline Inference