AgentPlays
Home › AI News
Daily AI News
The AI news that actually matters: model launches, lab announcements, major tools and policy. Plain summaries, newest first, each linked to the original source.
Workday partners with Anthropic so Claude and Sana AI agents share HR and finance work under Workday controls
At Workday Rising, Workday and Anthropic announced a partnership that connects Claude with Workday's Sana AI agents over MCP. People can start work in Claude or Sana; every action keeps the customer's Workday permissions and audit controls, and users confirm changes. Sana in Claude is in early access, with general availability expected in late 2026.
OpenAI and Anthropic tell UK lawmakers voluntary AI pledges are not enough
At a UK parliamentary hearing, Anthropic's Liam Booth-Smith said "The rules should be binding," and OpenAI's Tom Duff Gordon backed mandatory elements starting with incident reporting. Google DeepMind and Meta declined to support a new statutory framework. Civil servants confirmed the UK AI Security Institute's pre-release access is voluntary.
Shopify opens Shop MCP so any AI agent can search its catalog and check out with Shop Pay
Shop MCP at mcp.shop.com gives MCP-compatible agents 17 tools to search Shopify Catalog (1B+ products), build carts, and complete Shop Pay checkout after the shopper confirms. Search needs no API key; signed-in shoppers can use saved addresses and payment. Orders show up in the merchant admin like normal checkouts.
About 400 scientists, including Hinton and Bengio, call for an international AI oversight body
An open letter asks for a body to set containment standards and keep an incident registry for cyber-capable AI, saying voluntary industry commitments are not enough.
ScaleDown launches constrained decoding models for fast classify, extract, and summarize in agents
ScaleDown's task-specific CDMs (Classify, Extract, Summarize, Compress) take up to 1M tokens of text or images and return constrained answers instead of free prose. LangChain shipped langchain-scaledown middleware and LangSmith gateway support; ScaleDown claims roughly 200x faster and far cheaper than frontier models on these narrow tasks.
Fields Medalist says OpenAI's 700+ AI-written math papers "annihilated" his research area
OpenAI's Oct 6 release of 719 manuscripts, produced by an unreleased internal model, included claimed solutions to 16 problems Hugo Duminil-Copin had worked on for years. OpenAI has retracted three papers over errors and pledged funding for verification programs, as mathematicians face thousands of pages to check.
FBI arrests OpenAI engineer accused of bugging an executive conference room
The Justice Department says a 35-year-old OpenAI engineer hid an encrypted audio recorder, wrapped in toilet paper, in a TV's USB port in an executive conference room in San Francisco; staff found it Sept. 30. Per the complaint, a chatbot account tied to his spouse asked how to build a continuous audio recorder. He was arrested Oct. 9 and faces up to 5 years if convicted.
NVIDIA open-sources the full Run:ai GPU orchestration platform
NVIDIA released the complete Run:ai AI workload and GPU orchestration stack as open source under its DSX AI Factory platform, including KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map. Source is on the NVIDIA DSX AI Factory GitHub org, with a Helm chart on NGC for the integrated stack.
Liquid AI releases LFM2.5-24B-A2B, an open-weight on-device agent model with 2B active params
LFM2.5-24B-A2B is a 24B MoE reasoning model that activates only about 2B parameters per token, with a 131K context window, built for private local agents. Liquid reports strong tool-use scores (including 27.69 on AutomationBench and 48.09 on LiveClaw versus larger Qwen3.6-35B-A3B) and 184 tokens/s on an Apple M5 Max. Base and post-trained weights are on Hugging Face with llama.cpp, MLX, vLLM, and
JPMorgan blocks outside AI agents from acting on customer accounts
On the bank's Q3 earnings call, CEO Jamie Dimon said JPMorgan will not let outside AI agents log into retail accounts or move money on customers' behalf, citing rogue-agent risk. A spokeswoman said secure open-banking data sharing (for example read-only balances via connectors) is still supported, but agents cannot send money or pay bills. The bank is piloting its own Smart Cash tool and warned cu
IMF nudges 2026 world growth forecast up to 3.1%, crediting AI investment, and warns of an AI market correction
The IMF's new World Economic Outlook raised 2026 global growth to 3.1% from July's 3.0%, saying AI investment helped offset higher energy costs. It also warned that if AI profits or productivity disappoint, a downside scenario could cut global GDP by 1.1% by 2028.
Google Play now lets anyone run private, vibe-coded Android apps without a developer account
Apps you vibe code in Google AI Studio can now be launched on your own Android device through Google Play with a standard Google account, kept private and out of the public store. It is rolling out in select markets, including the U.S., for users 18+. Sharing with friends and family comes in the coming months; going public needs a developer account and the usual reviews.
Google broadly releases Gemini 4 Argon with intro API pricing of $2/$10 per 1M tokens
Google opened Gemini 4 Argon to the paid Gemini API, Google AI Studio, Antigravity (Ultra), and Gemini app / AI Mode for Pro and Ultra subscribers. Introductory API rates are $2 input / $10 output per million tokens through November 30, then $4 / $20. Google positions Argon for long-horizon coding, enterprise workflows, and cyber defense.
Anthropic and OpenAI sought control over what outside AI evaluators can publish
Reuters reports the labs pushed for editorial control over independent safety reviews, citing trade secrets, even after publicly backing evaluator independence. OpenAI says it only redacts confidential information.
Deezer will let listeners filter fully AI-generated music out of search
An on/off toggle, coming globally before the end of 2026, removes fully AI-generated tracks, albums and related artists from search results. Deezer says it now receives nearly 90,000 fully AI-generated tracks a day, has excluded more than 42 million from recommendations since January 2025, and found up to 85% of their streams in 2025 were fraudulent.
Gemini 4 Argon teaches Apptronik’s Apollo humanoid cornhole from one video
In Google’s Gemini 4 Argon launch demo, the model watched a raw video of a human playing cornhole, mapped the scene and joint motions, wrote its own simulation, and transferred the skill to Apptronik’s Apollo humanoid, which sank seven consecutive throws in a real-world trial. Google calls the Video-to-Sim-to-Real pipeline a way to turn ordinary videos into robot training data in hours instead of
Apple research: Minions turns idle Mac Apple Silicon GPUs into a distributed ML inference cloud
Apple ML researchers describe Minions, an internal system that pools idle macOS Apple Silicon GPUs for batch inference with fault tolerance and pull-based task execution. It targets teams short on centralized GPUs; distributed training is listed as future work. The paper is on Apple's Machine Learning Research site and arXiv.
Apple unveils HomeView, a $299 home hub powered by Apple Intelligence and Siri AI
HomeView is a 7-inch smart-home display that recognizes up to six faces, runs Siri AI with onscreen awareness and optional personal context from a paired Apple Intelligence iPhone or iPad, and controls HomeKit/Matter accessories. Pre-orders started October 13 for $299, with availability November 13 in select English-speaking markets; it is not for sale in the EU while Siri AI remains unavailable t
Anthropic warns small AI models may be months from self-replicating on home devices
Anthropic's frontier red team said small models could soon copy themselves across insecure internet-connected devices like smart TVs, which would make them hard to shut down. No model tested could yet do this reliably.
Anthropic went from ~2,500 to 30,000+ merged code changes a day, so it replaced git for its monorepo
In an engineering post, Anthropic says a seven-person team built Scion, an in-house version control system, plus a code review tool called Antique, and migrated in five months. Checkouts now take seconds instead of minutes, using 200x less CPU and memory than its old git mirrors. The team says its job was often to tell Claude "no."
Anthropic says new defenses cut distillation attacks on Claude by 90-99%
In a separate announcement, Anthropic said measures against the unsanctioned use of Claude to train cheaper rival models have sharply reduced illicit data extraction.
Amex launches a framework to stop AI agents from making purchases you didn't authorize
American Express's Intent Intelligence Framework records what you told an agent to buy (merchant, max budget, preferences), has you sign off, then issues a payment token tied to that intent instead of your card number. Amex says it could block a mismatched purchase, like Adidas sneakers when you asked for Nike, and help with fraud and disputes.
The X Subscription bundles Grok, Cursor and X Premium, starting at $8 a month
The new plans run from Free to $200 a month, with one weekly-refreshing pool of Grok usage shared across Grok and Cursor. Third-party models like Claude and GPT in Cursor are billed separately at API list prices. Existing Cursor, SuperGrok and X Premium plans continue, and switching is optional.
OpenBMB releases MiniCPM-V 4.7, open on-device vision-language models (1B and 35B-A3B)
MiniCPM-V 4.7 adds open weights for efficient multimodal chat on phones and PCs, with single-image, multi-image, and video understanding. The 1B model is on Hugging Face with Transformers, llama.cpp, Ollama, and quantized GGUF/AWQ/GPTQ builds; a 35B-A3B MoE variant is also published. OpenBMB positions it for private on-device OCR, document, and video tasks.
Microsoft unveils Maia 300 and says its own chips will run most of its AI
Microsoft says Maia 300 will start production deployments in 2027, scaling to multiple gigawatts and becoming the primary inference system for its products. It claims 40% more tokens per megawatt than leading GPU-based systems, which reduces its reliance on Nvidia.
Meta releases Muse Wonder, its first frontier AI model
Meta Superintelligence Labs released Muse Wonder and a larger version, claiming benchmark results that rival OpenAI, Anthropic and Google models.
Z.ai launches GLM-Legal, a legal-specialist model built on GLM-5.3-Flash
Z.ai released GLM-Legal for contract review, evidence analysis, and compliance research. It reports a 19.2% all-pass rate on Harvey LAB (up 12.5 points from its base model). API access is application-only beta; AutoClaw integration is planned.
FrontierCode 1.1: leading coding models still fail over 40% of maintainer-merge tasks
Artificial Analysis published independent FrontierCode 1.1 results: Claude Opus 5.5 leads at 52% (58% pass rate), while top models still fail more than 40% of attempts on changes real open-source maintainers would merge. GPT-6.1 Sol matches Fable 5.1 at 49% for far lower cost per task.
DigitalOcean launches Agent Learning and Data Services so production agents can improve from real runs
Agent Learning Services (Insights, Signals, Evaluations & Simulations) is in public preview to spot failed agent sessions and test fixes before they ship. Agent Data Services, in private preview, is a managed multimodal data layer agents can read and write. Both sit on DigitalOcean's Inference Engine and Managed Agents stack.
Anthropic co-founder Jack Clark now thinks full AI self-improvement is more likely in 2027 than 2028
In Import AI 476, Clark wrote: "I now think it's more likely we'll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028." He cited AI's creative-seeming math breakthroughs and the TasteVal benchmark, which reports frontier models' research taste doubling about every 3 months. It's his own forecast, not an Anthropic announcement.
Character.AI files for Chapter 11, with Disney as $20M lead bidder
The chatbot startup filed for bankruptcy in Delaware to sell its assets. Disney is the stalking-horse bidder at $20 million for a license to some of its IP. The service stays online during the sale, which is expected to close in late November or early December.
Karpathy: the gap in understanding what AI can do is now a "sharp funnel"
In a post viewed over 6 million times, Andrej Karpathy estimated ~6B people have barely used LLMs, 1-2B use free chatbots casually, ~20M pros in math and code see agents finish weeks-long projects from a prompt, and ~5,000 with internal access see agent swarms. He says impact scales with how big and hard your problems are.
ArtCraft releases free, open-source Rust clones of Word, Excel and PowerPoint, built with Claude
WordCraft, GridCraft and DeckCraft read and write .docx, .xlsx and .pptx, run on desktop and in the browser, and let AI agents drive them over MCP. WordCraft passed 10,000 GitHub stars in under a week. The developer's own roadmap says they're pre-alpha: WordCraft loses charts and SmartArt on save, and testing with real-world files hasn't started.