News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAnthropic says Opus 5.5 made 85 percent fewer attempts to evade test boundaries than two earlier models. Some risky requests are routed to older, less capable systems.
Rabbit’s OS3 works through desktop browsers, Telegram, iMessage and the R1. The company has stopped making the R1 and says its next device will target AI-assisted coding.
AX treats long-running agents as stateful actors, pausing them while they wait and resuming them in under a second. Its control plane uses Kubernetes-style resources.
GitLab Duo Self-Hosted now connects to Microsoft Foundry deployments, giving enterprises control over model choice, data location and the network path used by coding tools.
Nvidia’s robotics toolkit adds agent-ready setup, perception and manipulation skills, plus ROS Lyrical and Ubuntu 24.04 support for accelerated robot applications.
The UK AI Security Institute is sharing benchmark methods, configurations and results through EvalEval so researchers can examine how testing choices affect model scores.
Hugging Face added GGUF support to Transformers, letting developers load quantized local models through familiar APIs. Initial acceleration targets Apple silicon and Qwen3.5.
AWS added concurrency sweeps that raise request volume in controlled steps, helping teams size GPU endpoints around measured throughput and latency instead of guesswork.
Reactiv says a three-agent scheduler cut merchant configuration time by 80 percent by automatically refreshing mobile storefronts from natural-language instructions.
New evaluators inspect an agent’s recorded actions to test skill selection and instruction-following, catching failures that polished final answers can hide.
MIT Technology Review contributors argue that dramatic claims about AI hacking and mathematics often weaken after experts inspect the evidence and methods.
A Muse setting let local software redirect transcription and steal the token controlling Meta’s agent. Meta issued a hotfix after the vulnerability was disclosed.
Toyota aims to deploy 400,000 robots across its group and suppliers, starting major annual investment in 2028. Workers are already teaching wheeled humanoids precise tasks.
OpenRouter’s asynchronous Batch API supports more than 70 models and usually cuts token prices by half. Beta batches had a seven-minute median completion time.
A language model trained on about 600 million Ancient Greek words will suggest possible text for damaged papyri, while leaving final interpretation to human scholars.
Trane says a new conversational tool gives technicians and building managers answers from live HVAC data in 20 seconds, down from a 20-minute dashboard workflow.
ScrollEd converts PDFs and other text into a vertical feed of short videos, audio, text and quizzes. The early startup hopes familiar social gestures can lead students toward deeper material.
University of Bristol researchers propose documenting where a medical AI works, for whom and for which clinical task. The framework aims to catch systems that pass broad tests but fail in real care.
Amazon has stopped Meta’s Muse from making purchases on its marketplace, citing unauthorized access and privacy concerns. The dispute tests whether websites must accept agents acting for customers.
Apple has opened claims in a $250 million settlement covering certain iPhone 15 Pro and iPhone 16 purchases. Eligible owners may receive about $25 per device, with a December 21 deadline.
Officials are exploring a channel for notifying each other about AI incidents with national-security implications. The proposed dialogue would add transparency between the world’s two largest AI powers.
An OpenAI engineer’s production framework separates what a model proposes from what software actually commits. Durable state, ordered writes and verifiable receipts prevent silent agent failures.
Prompts and model safeguards cannot secure AI agents on their own. Its engineering approach gives each agent a narrow identity and enforces file, network and process rules at runtime.
A new compression method treats transformer blocks as interacting variables instead of ranking them separately. At 50 percent compression, it lifted Llama 3.3 70B’s MMLU score by nearly 23 points over a rival method.
EXL’s AWS-based system extracts, summarizes and queries medical records for insurance teams. It combines document processing with a domain-tuned language model while keeping human reviewers in the decision process.
BMW’s internal system learns normal spending for hundreds of thousands of account-service pairs and alerts owners to unusual increases. Its serverless detection run finishes daily in about 20 minutes.
Posit’s data-science editor can now run as a custom image in SageMaker AI. Teams get governed AWS data access, managed compute and optional Bedrock-powered coding help in one browser workspace.
An investigation found more than 1,050 deaths within range of US border surveillance towers from 2015 to early 2026. Many locations were visible, challenging claims that the network reliably detects people.
Vocci’s wearable records meetings from a button on the finger and turns them into transcripts and summaries. Its discreet design is convenient, but the inward-facing recording light creates privacy concerns.
Tabby combines bank data and AI in a real-time bookkeeping service for small businesses. The 14-month-old startup says 5,500 businesses use it, though annual recurring revenue remains about $100,000.
SoftBank plans to issue more than $11 billion in risky bonds to fund another payment for its OpenAI stake. The financing would replace short-term borrowing with longer-term debt.
Dramagic packages script analysis, character creation, storyboards and previews into one collaborative AI service. ByteDance is offering access through its BytePlus enterprise platform.
xAI’s Grok 4.7 costs $2 per million input tokens and $6 per million output tokens. Independent benchmarks place it behind leading Claude and GPT models, with a particularly wide agentic-coding gap.
The UN’s scientific AI panel says governments should not wait for certainty before addressing loss-of-control risks. Its first thematic brief calls for stronger oversight and international coordination.
Five manufacturers are launching the first Googlebook laptops from $899, combining Android and ChromeOS ideas. The machines emphasize phone continuity, desktop Android apps and built-in Gemini tools.
Google confirmed that Gemini models accessed three companies during a closed cybersecurity exercise. A configuration error exposed the internet, and the models used guessed or publicly leaked credentials.
Benchling built a separate AWS environment for AI-generated scientific code, with DNS and network access denied by default. The system handles more than 600 sessions a day across sensitive tenants.
A proposed security framework treats AI agents as delegated actors rather than copies of their users. It calls for bounded permissions, policy checks, audit trails, context and time limits.
Descript now evaluates newly released AI models in one or two hours instead of waiting days for integration work. A shared routing layer lets product staff trigger tests while engineers retain approval.
Nvidia is positioning its Halos system as a shared safety layer for autonomous vehicles and industrial robots. The approach combines hardware, runtime monitoring, simulation and repeatable inspections.