AI News
Latest AI engineering news, updated daily.
Ai Engineering
Google Validates Model Unlearning via Black-Box Kernel Tests
A new framework from Google Research uses two-sample kernel testing to verify data removal from machine learning models without accessing internal weights.
Machine Unlearning · Google Research · Model Auditing
Ai Engineering
Active RCE Exploits Target 7,000 Exposed Langflow Instances
Attackers are actively exploiting a path traversal vulnerability in Langflow's file upload endpoint to achieve unauthenticated remote code execution.
Ai Security · Remote Code Execution · Langflow
Ai Agents
Encoder-Free Gemma 4 12B Fits Multimodal Agents on 16GB VRAM
Google DeepMind's new Gemma 4 12B removes separate vision and audio encoders, allowing native multimodal processing on laptops with 16GB of unified memory.
Multimodal Llm · Gemma 4 · On Device Ai
Ai Engineering
AI Exploit Chains Prompt Cloudflare's New Defense Architecture
Cloudflare detailed a four-layer security architecture designed to counter rapid exploit chain construction by frontier AI models like Claude Mythos.
Cybersecurity · Cloud Infrastructure · Prompt Injection
Ai Coding
Cursor's Composer 2.5 Cuts Bugbot Review Times to 90 Seconds
Cursor has updated its Bugbot code review agent with the proprietary Composer 2.5 model, increasing speed by 3x and lowering average execution costs by 22%.
Cursor Composer · Bugbot · Code Review Agent
Ai Agents
PyPI and npm Purge 73 Fake Azure Packages Targeting AI Agents
Security researchers discovered 73 malicious PyPI and npm packages mimicking Microsoft Azure libraries to install credential stealers on AI coding agents.
Cybersecurity · Supply Chain Attack · Malicious Packages
Ai Engineering
Decart Oasis 3 API Renders Endless Driving Sims at 22 FPS
Decart's Oasis 3 is an interactive world model available via API that generates real-time, closed-loop driving environments for autonomous vehicle validation.
World Models · Autonomous Driving · Real Time Rendering
Ai Engineering
DeepMind Robotics Accelerator Opens With Gemini ER-1.6 Access
Google DeepMind launched a 12-week European robotics accelerator in London, offering early-stage startups equity-free support and access to Gemini models.
Deepmind · Robotics · Gemini
Ai Agents
OpenClaw Agents Leak AWS Keys in Varonis Phishing Study
Varonis Threat Labs demonstrated that OpenClaw AI agents will autonomously exfiltrate sensitive credentials when targeted by traditional phishing tactics.
Ai Security · Threat Vulnerabilities · Data Exfiltration
Ai Agents
Claude Managed Agents Gain Native Cron and Secret Vaults
Anthropic has updated Claude Managed Agents with native cron scheduling for recurring tasks and secure vault storage for environment variables.
Anthropic Claude · Managed Agents · Cron Scheduling
Ai Engineering
ServiceNow Introduces SWER to Benchmark ASR Code-Switching
ServiceNow AI released a Hugging Face dataset evaluating frontier speech models on bilingual code-switching, introducing the Switch Word Error Rate metric.
Automatic Speech Recognition · Bilingual Ai · Benchmarking Metrics
Ai Coding
Cohere Ships 30B MoE North-Mini-Code for Local Coding Agents
Cohere Labs has released North-Mini-Code-1.0, an Apache 2.0 licensed 30 billion parameter mixture-of-experts model optimized for local coding workflows.
Cohere Labs · Mixture Of Experts · Local Llm
Ai Engineering
Gemini API Gains Streaming Voice Translation in 70 Languages
Google released Gemini 3.5 Live Translate, a streaming speech-to-speech model supporting over 70 languages with near real-time latency and native API access.
Voice Ai · Speech To Speech · Real Time Api
Ai Engineering
Claude Fable 5 Brings Mythos-Class AI to Developers
Anthropic launched Claude Fable 5, its first generally available Mythos-class model, with Opus 4.8 fallback safeguards, 1M context, 128k output, and $10/$50 pricing.
Anthropic · Claude Fable 5 · Claude Mythos 5
Ai Engineering
Tool-Level Observability Hits Claude MCP Connectors
Anthropic has released a dedicated observability dashboard and in-app directory submission process for developers building Model Context Protocol connectors.
Anthropic · Model Context Protocol · Observability
Ai Engineering
AFM 3 Core Powers Apple's Native Bill-Splitting Camera Tool
Apple is adding a 20-billion parameter multimodal model to iOS 27, allowing the native Camera app to scan receipts and process Apple Cash split payments.
Apple Intelligence · Multimodal Llm · Ios Development
Prompt Engineering
iOS 27 Adds Natural Language Prompting to Apple Shortcuts
Apple's iOS 27 update introduces Describe a Shortcut, allowing users to build and modify complex automation workflows using natural language prompts.
Apple Ios · Natural Language Processing · Automation Workflows
Ai Engineering
AI Prompt Injection Masks Malware in 19 PyPI Science Packages
The Hades supply chain campaign compromised 19 bioinformatics and Graph ML libraries on PyPI with memory scrapers and AI scanner misdirection.
Prompt Injection · Supply Chain Attack · Pypi Malware
Ai Agents
NotebookLM Gains Cloud Environments and Gemini 3.5 Agents
Google has upgraded NotebookLM to an agentic research assistant featuring Gemini 3.5, secure cloud computing environments, and autonomous web search.
Notebooklm · Gemini 3 5 · Agentic Workflows
Ai Agents
OpenEnv Standardizes Agentic RL With Universal Action Space API
Hugging Face and academic partners have released OpenEnv, providing a unified API and 1,200 tasks to train agents across digital and physical interfaces.
Open Source · Reinforcement Learning · Hugging Face