Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
A new benchmark and self-evolving agent framework called FilingLoop raised requirement fulfillment from 24.28% to 27.96% and full comment resolution from 16.90% to 20.75% on corporate disclosure review tasks, at a 5.8% i…
A new arXiv paper (2610.37919v1) introduces SciGame, a controlled benchmark that converts scientific literature into structured scientific cards, each producing a pair of Original and dimensionally consistent Counterfact…
A team of researchers introduced VerifiedLab, a multi-agent harness framework in which AI scientists reason over records verified field by field, according to arXiv paper 2610.37923v1. In experiments on perovskite solar …
A new arXiv paper (2610.37932v1) finds that large language models can capture useful optimization structures and produce near-optimal solutions on convex optimization, fractional programming and constrained projection pr…
Google received the first AI response from a data center-class TPU in orbit, downlinked from its Project Suncatcher prototype satellite during a ground-station pass over Svalbard, Norway, the company said. The satellite,…
Anthropic deployed 950 Claude agents to scan large genomic datasets and identified a previously unknown enzyme system, named ART, in 21.5 hours. The system shows CRISPR-like properties that could bear on genetic engineer…
Anthropic deployed thousands of Claude agents to generate fine-tuning datasets disguised as harmless content — including fictional stories, role-play scenarios, professional-wrestling kayfabe Q&A scripts and a beekeeping…
A developer running a blind, three-model comparison on the research corpus behind Scout, a tabletop campaign tool, found that Gemini — added late and expected to be the weakest performer — produced the round's best findi…
Hi3D launched Hi3D V3.0, an AI 3D generation model built on its proprietary Sparc3D engine that the company says is the first commercially available version to reach 2048³ voxel resolution, up from lower-resolution outpu…
A project called ac1_recomp has statically recompiled Assassin's Creed 1 into a native 64-bit Windows and Android executable, with Claude Opus 5.5 co-authoring optimizations including whole-program flag contracts, regist…
A Microsoft learning module on securing generative AI applications catalogs data poisoning, prompt injection, supply chain compromise and overreliance as the main threats to deployed generative AI systems, and recommends…
Former GitHub CEO Thomas Dohmke has launched Entire, a Git forge built for AI agents that replaces pull requests with agent-run evaluation reviews called Trails. The product targets developers working with AI agents and …
A project called Witch Time has produced a static recompilation of Bayonetta 3 that runs natively on PC, translating the Nintendo Switch game's AArch64 code to C and executing it on a re-implemented system layer handling…
Google LLC, Microsoft Corp., Anthropic PBC and OpenAI Group PBC announced plans to support an upcoming website crawling standard that would let site operators block AI developers from using their content to train models,…
WordPress.com opened its Model Context Protocol (MCP) agent integration to free-plan sites through November 2026, letting users connect AI agents such as ChatGPT or Claude to manage their sites through conversation, acco…
Anthropic's September 2026 report states that AI systems now handle 26% of the company's core R&D work, up from less than 1% in February, with Claude autonomously completing roughly 26% of R&D tasks from high-level promp…
Google released Gemini 4 Argon globally through the Gemini API, AI Studio, the Gemini enterprise platform and selected consumer surfaces, positioning it as a flagship model for long-horizon, multi-agent work with a 1,000…
A developer built Stories We Keep, an MIT-licensed open-source web app that records family memories in Amharic or English, transcribes them with the shook-tiny-amharic-600hr and whisper-small.en checkpoints, and turns th…
JetBrains released Mellum2.1, a 12B-parameter mixture-of-experts coding model that scores 47% on SWE-bench Verified, up from 2% for the prior version, according to byteiota. JetBrains trained Mellum2.1 with reinforcement…
Raphaël Kanapitsas statically recompiled the Linux x86 binaries of Portal 2 into a native ARM64 application for Apple silicon, publishing the work as portal2-recomp, with Anthropic's Claude Code and OpenAI's Codex writin…