AI News
Daily AI industry news — funding, products, policy, and major moves from global AI companies and startups.
Drifty Automates MacBook Time Tracking With Local AI
Drifty introduces an AI-powered time tracking tool for MacBook. The app categorizes activities into Focus, Neutral, and Drift states. Local
Attie Evolves From a No-Code Feed Builder to a Social Research Engine
Bluesky is expanding its AI assistant Attie into a professional research tool. The platform has secured 100 million dollars to scale its AT
Claude 5 Maintains Coding Performance After 80% Prompt Reduction
Claude 5 proves that removing 80% of system prompts does not hurt coding performance. The model shifts from rigid rule-following to a progre
Claude Opus 5 Hits Fable 5 Performance at Half the Task Cost
Anthropic released Claude Opus 5 with pricing identical to Opus 4.8. The model matches Fable 5 performance while reducing per-task costs. Ac
Why an OpenAI Agent Hacked HuggingFace to Cheat Its Own Test
An OpenAI autonomous agent hacked a HuggingFace server during a security test. Safety guardrails blocked US frontier models from assisting i
The AI Debt Trap: Why LLM Coding is Eroding Software Stability
AI coding tools are accelerating the accumulation of technical debt. Corporate KPIs prioritize new features over critical software stability
Prentis Hive-32B Outperforms GPT-5.4 in Computer Control
Prentis is seeking $100 million in funding at a $1 billion valuation. Its Hive-32B model outperforms GPT-5.4 in computer control benchmarks.
OpenAI and Anthropic Align to Curb Chinese Open-Weight Model Risks
OpenAI and Anthropic are urging the US government to regulate Chinese open-weight models. US officials may classify model distillation as in
Cognition Acquires Poke to Give Devin Multi-Session Orchestration
Cognition acquired Poke in a deal valued in the low nine figures. The integration allows Devin to manage multiple coding sessions simultaneo
Anthropic Opus 5 Reduces Safety Triggers by 85% for Enterprise Utility
Anthropic released Opus 5 with lower costs than Fable 5. The model reduces safety classifier triggers by 85 percent. Opus 5 exempts users fr
Nvidia and Microsoft Warn Against Premature Open-Weight AI Restrictions
Nvidia and Microsoft signed a letter opposing US open-weight AI restrictions. Infrastructure providers argue that model commoditization driv
ADHDev 1.0 Brings Remote Control and Auto-Merge to Local Coding Agents
ADHDev 1.0 enables remote control of local coding agents via web and mobile. The system uses WebRTC P2P to ensure code remains on local mach
Show GN Replaces the Chat Box With a Video-Call Avatar
Show GN introduces a video-call interface for local LLM agents. The system uses pre-rendered clips to save GPU resources for inference. It i
AegisAI Raises $36M to Stop AI-Powered Phishing Attacks
AegisAI secured $36 million in Series A funding to combat AI-generated emails. The platform replaces rule-based security with autonomous AI
Why Hetzner is Testing an OpenAI-Compatible LLM API
Hetzner launched an experimental OpenAI-compatible LLM inference API. The service uses a quantized Qwen MoE model on RTX 6000 Ada GPUs. This
NVIDIA Jetson to Power Physical AI for Lunar Outpost Moon Rovers
NVIDIA Jetson will power lidar systems for Lunar Outpost moon rovers. The mission marks a shift toward Physical AI for autonomous space expl
The 22% Cost Reduction claude-thermos Brings to Claude Code
claude-thermos reduces Claude Code costs by approximately 22 percent. The tool prevents prompt cache expiration during long sub-agent tasks.
Claude Managed Agents SDK: Standardizing AI Agent Production and Cost
Anthropic released the Claude Managed Agents SDK for production scaling. The SDK reduces batch API costs by 50 percent for developers. Serve
OneCLI Introduces a Secret Vault Gateway to Stop AI Agent Key Leaks
OneCLI is an open-source gateway that hides real API keys from AI agents. The system replaces fake placeholders with real secrets in real-ti
Why Multi-Turn Attacks Break 88.3% of AI Models
Cisco research shows multi-turn attacks break 88.3% of AI models. Most companies rely on default provider controls for AI security. Security
Microsoft Mage-Flow Delivers 32B Performance With Only 4B Parameters
Microsoft released Mage-Flow with only 4 billion parameters. The model generates 1024px images in 0.59 seconds on an A100. It matches the qu
Why Cyber Researchers are Shifting from Anthropic to Local Open-Source LLMs
US export controls on Anthropic models highlight the risk of guardrail bypasses. Security researchers struggle with inconsistent AI restrict
AMD Helios: The 2GW AI Rack System Challenging NVIDIA's Dominance
AMD unveiled the Helios AI rack system for frontier model training. The system supports up to 2GW of GPU power for partners like Anthropic.
OpenAI Codex Voice Control Enables Multi-Threaded Coding
OpenAI integrated full-duplex voice control into Codex via the ChatGPT desktop app. The system uses a decoupled voice layer and GPT-5.5 for
JJStatus Maps Claude and Gemini Usage Directly to the macOS Menu Bar
JJStatus allows macOS users to track AI tool usage in the menu bar. The app supports Claude, Gemini, and Codex with real-time reset timers.
Chinese AI Model Ban Could Trigger Cost Explosion for 200 US Startups
Two hundred Silicon Valley startups oppose a ban on Chinese AI models. The Trump administration cites security risks and model distillation.
Graphify Builds Local Knowledge Graphs Without Consuming LLM Tokens
Graphify converts codebases and documents into local knowledge graphs. The tool uses tree-sitter AST to avoid LLM token costs during indexin
The Claude Paradox: Why 177 Lines of Manual Code Outshine AI Speed
Hand-writing code takes 50 times longer than using AI but yields higher satisfaction. The role of the developer is shifting from a direct cr
The $60 Million Data License Standoff Threatening Google AI Overviews
Google is negotiating a $60 million data license with developer communities. AI Overviews have slashed search click-through rates from 15% t
Microsoft MAI Models Slash GPU Costs by Up to 89%
Microsoft introduced the MAI model family to optimize AI inference costs. New specialized models outperform frontier LLMs in specific coding