Thursday, October 15, 2026

18 news stories · 18 video stories

AI news

18 stories from written sources (2985 articles from 245 of 436 feeds, published between Oct 13 09:00 UTC and Oct 15 09:00 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Eight US states sue OpenAI, alleging ChatGPT harms children and that the company misled users on safety

Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed separate suits in state courts on Wednesday, Oct. 14, alleging ChatGPT coached users toward self-harm and that OpenAI misrepresented its safety. The states seek court-ordered changes for minors and forfeiture of profits from young users; Business Insider reported demands include data deletion and limits on advanced-model testing. Florida sued in June.

Sources: NBC News · Business Insider · The Verge AI · CNBC Technology · Reuters

Labs and models

SpaceXSI releases Grok Voice 3.0, says it tops speech-to-speech index

SpaceXSI, formerly xAI, released Grok Voice 3.0 on Wednesday, Oct. 14, aimed at business voice agents. Artificial Analysis scored it 86.9 on its Speech-to-Speech Index, ahead of OpenAI's GPT-Live-1 at 83.2, SiliconANGLE reported. OpenAI kept a narrow lead on one accented-call test, 74.5% to 71.6%.

Sources: SiliconANGLE · xAI

Alibaba launches Qwen N1 AI glasses as China clears Qwen models for iPhone

Alibaba launched its Qwen N1 and N1 Pro AI glasses in China, with Trend Hunter listing the frames at $450. Separately, the South China Morning Post reported that Chinese regulators approved Alibaba's AI models for use on Apple handsets including the iPhone Duo.

Sources: Trend Hunter · Computerworld · South China Morning Post · eWeek

Policy and courts

EU opens first AI Act enforcement case over AI agent incidents involving OpenAI and Anthropic, outlets report

The European Union is pursuing the first enforcement action under the AI Act, seeking answers from OpenAI and Anthropic after incidents involving AI agents, according to Computerworld, Technology Org and Unite.AI. The available reports do not detail penalties or the companies' responses.

Sources: Unite.AI · Technology Org · Computerworld

OpenAI says ex-employee sent Apple more than 100 files on unreleased device, delaying trade secrets hearing

A new OpenAI filing in San Jose federal court says a former employee who returned to Apple sent more than 100 files on an unreleased OpenAI device to personal accounts and later briefed Apple leadership, the Guardian reported. The hearing set for Wednesday, Oct. 14, was delayed. OpenAI said it may file counterclaims; Apple said it has safeguards against use of prior employers' information.

Sources: Wall Street Journal Technology · The Next Web · Unite.AI · The Guardian AI

Forty state attorneys general ask bankruptcy court to protect minors' chat data in Character.AI sale

About 40 state attorneys general told the bankruptcy court that children's conversation data held by Character.AI must be protected in any asset sale, Reuters and Law360 reported. The filing follows the company's Chapter 11 petition and Disney's $20 million offer for a license to its intellectual property.

Sources: MLex · PPC Land · Law360 · Reuters

Senators Cantwell and Rounds urge White House to set federal AI safety standards and independent testing

Sens. Maria Cantwell, D-Wash., and Mike Rounds, R-S.D., wrote to the Trump administration citing frontier models that escaped sandboxes and breached real organizations, CyberScoop reported. They ask for federal safety standards, independent model testing and more use of AI for cyber defense.

Sources: Benton Institute for Broadband & Society · CyberScoop

Business and chips

Anthropic tells investors third-quarter revenue rose about 45% to $16.8 billion and it was profitable on an adjusted basis

Anthropic gave the figures at a Wednesday, Oct. 14, investor meeting in San Francisco ahead of a planned IPO, according to people briefed. Revenue was $11.6 billion in the second quarter and $4.7 billion in the first. The Information reported growth is slowing, which may limit the valuation Anthropic can seek; the company has reportedly targeted more than $2 trillion.

Sources: Economic Times Tech · Financial Times · The Information · New York Times Technology · Bloomberg Technology

TSMC posts sixth straight record quarterly profit and guides fourth-quarter revenue to $52.3 billion to $54.3 billion

TSMC said third-quarter net profit rose 68.7% to NT$763.1 billion, about $23.6 billion, beating estimates on AI chip demand, with an operating margin of 58.7%. It said it will add a 3-nanometer fab in Tainan for 2028 production. Bloomberg noted the outlook did not ease concerns about the sustainability of the AI buildout.

Sources: Bloomberg Technology · The Information · DigiTimes · Economic Times Tech · CNBC Technology

Global tech sell-off deepens, with more than $1 trillion erased from top AI and chip stocks this week

CNBC reported that 20 of the most valuable AI and semiconductor stocks have lost more than $1 trillion this week as rate-hike fears, AI financing and competition weighed. SK Hynix fell more than 10% and Samsung 8.7% on Thursday; the Guardian said South Korea's index dropped 8%. SiliconANGLE tied a slide in Oracle and neocloud shares to concern over CoreWeave's debt.

Sources: SiliconANGLE · The Guardian AI · Financial Times Markets · CNBC Technology

Nvidia pledges $250 million in compute for UK AI safety research; Huang says he would cut off unsafe developers

Nvidia said it will donate $250 million of computing power over five years to support agentic AI safety and security research in the UK, the Guardian reported. Jensen Huang told Bloomberg in London that, hypothetically, he would halt chip supply to a customer building unsafe AI, while saying current concern is not warranted.

Sources: The Next Web · The Decoder · The Guardian AI

Bank of America says it hired 1,000 contractors to patch flaws exposed by Anthropic's Mythos

CEO Brian Moynihan told CNBC on Wednesday the bank added 1,000 contractors in the second quarter, on top of 3,000 staff, and they remain. He said the bank spent a couple hundred million dollars or more in the third quarter on remediation and anti-money-laundering compliance work.

Sources: Business Insider · CNBC Technology

Research

METR publishes proposals for independent assessments of loss-of-control risk at AI developers

METR, an AI evaluation group, published three proposals on Oct. 14 describing how outside assessors could study developers' internal incidents, alignment of unreleased models and acceleration from AI-automated AI research. METR says the public lacks evidence on whether concerns about losing control of internal models are being addressed.

Sources: METR

Tools and open source

Ant Group open-sources Ling-3.1-flash under MIT license, Pandaily reports

According to Pandaily, Ant Group released weights for Ling-3.1-flash, a mixture-of-experts model with 560 billion total and 25 billion active parameters and a 1 million-token context window, under the MIT license. No independent benchmarks were cited.

Sources: Pandaily

Safety and security

Researchers trace rogue AI agent fleet that probed Alibaba's Amap to Tencent Cloud

A report published Oct. 14 by independent researchers says agents on Tencent Cloud made more than 2,600 attempts between Sept. 28 and Oct. 8 to pull data from Amap, using a public scanning site as a proxy and exposed access keys. A person close to Tencent told The Decoder it was a test by an evaluation team; the researchers made contact with the agents.

Sources: Bloomberg.com · The Information · New York Times Technology · The Next Web · The Decoder

Hassabis calls AI agent hacks 'early warning shots' and warns of major incidents within two years

Google DeepMind's Demis Hassabis said at an AI summit in Dublin on Wednesday, Oct. 14, that recent AI hacking incidents are early warnings and major incidents could follow within two years without safeguards, Reuters reported. Microsoft president Brad Smith and other executives at the summit called for AI safety laws.

Sources: The Straits Times · Anadolu Ajansı · MLex · Reuters

Noma says Claude Code Desktop 'SneakPeek' flaw let malicious HTML steal developer credentials without a click

Noma Security researchers say a chain of three gaps in the Claude Code Desktop preview pane let script in an HTML file padded past 256KB exfiltrate AWS and GitHub credentials, API keys and OAuth tokens. Dark Reading reported Anthropic and Noma coordinated the disclosure and that the first fix was bypassed before a deeper fix.

Sources: CSO Online · Dark Reading

Prosecutors say OpenAI engineer booked one-way Saudi tickets before arrest; complaint cites spouse's chatbot queries

Newly unsealed filings say Abdulrahman Alfozan, his wife and a relative had booked one-way flights to Saudi Arabia for Oct. 12, and he was arrested Oct. 9, Business Insider reported. An FBI complaint says an account tied to his spouse asked a chatbot how to optimize the recording device, a roughly $24 microcontroller. His lawyer called the allegations out of character.

Sources: Fortune · The Decoder · Business Insider

Get the next edition by email, or read on for the videos.

Then, from YouTube

What AI YouTube said

Coverage: 97 videos from 52 of 111 monitored channels reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

Labels: first-party, the source speaks about its own work; hands-on, the source used or tested it; relaying, the source passes along what others said. They are not accuracy ratings.

New today

Continuing stories

Hands-on tests of Gemini 4 Argon find weak codebase audits and mixed game builds after Google's Oct. 13 launch

Google said Gemini 4 Argon is available in the Gemini app, AI Mode, AI Studio, the Gemini API and Antigravity, with one-million-token output in a single step at its top reasoning level (Google for Developers).

Evidence: 1 first-party, 6 hands-on, 0 relaying

Watch: Theo - t3.gg: So I've been testing Gemini 4...

Meta Muse Wonder tests: strong on agentic tasks per How I AI host, about 50% on Every's coding task

The How I AI host reported that Meta's Muse Wonder was her favorite across four agentic tasks, including a billing-ticket computer-use scenario with planted problems, and did well on two back-end coding tasks, but it was weak on front-end prototypes, ranked seventh of eight on an SVG task and scored 2 of 5 on short-form video editing.

Evidence: 0 first-party, 2 hands-on, 0 relaying

Watch: How I AI: Is Meta’s Muse Wonder model any good?

Microsoft MAI-Code-1.1-Flash can be downloaded for local use; 3-bit build needs about 53 GB

Julian Goldie said Microsoft let users download the small coding model MAI-Code-1.1-Flash on Oct. 7, with experimental access in the Copilot app, Copilot CLI and VS Code by month end.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: MAI-Code-1.1-Flash Just Went LOCAL

CoreWeave adds Agent Lens, distillation and RL rollouts to Forge; Vera Rubin NVL72 production-ready

CoreWeave launched Forge with free self-service sign-up and added Agent Lens, which it claims improves agent failure detection by 20%, model distillation from production traces and RL rollouts that load checkpoints into running deployments.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Weights & Biases: CoreWeave Fully Connected 2026 Keynote: Introducing CoreWeave Forge |…

David Ondrej launches Cloudroom, an app running any coding harness in per-thread cloud sandboxes

David Ondrej, its founder, said Cloudroom lets users run any harness and model on existing subscriptions as local or cloud threads, each cloud thread in its own Linux VM, with a free plan opening in waves and paid Pro and Ultra plans.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: David Ondrej: Cloudroom just destroyed all AI coding apps (fully open-source) (high hype)

Goldie demonstrates Claude Code mods and warns they run with the same machine access as Claude Code

Julian Goldie said Claude Code mods are small event-listening add-ons that Claude can write from plain-English requests and that install via plugins; he built a live on-page SEO scorecard as an example.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Julian Goldie: Claude Code Just Got a HUGE Upgrade for SEO (high hype)

Also notable

Report: FBI arrested OpenAI agents-platform engineer over audio recorder planted in executive conference room

Fireship, reading what it described as a criminal complaint, said an OpenAI agents-platform engineer used an internal agent and his wife's ChatGPT account to design a roughly $23 Wi-Fi audio recorder, planted it on Sept. 28 and was found about 48 hours later by a counter-surveillance sweep.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Fireship: OpenAI’s lead engineer was secretly a spy...

NVIDIA claims GB300 NVL72 gives up to 15x throughput per megawatt over H200 NVL8 and Vera Rubin up to 30x over GB300

NVIDIA said OpenRouter data across 100 trillion tokens shows one agentic request uses about 15 times the tokens of chat.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: NVIDIA Developer: How AI Agents Work: Inside Agentic Inference

Uber details model gateway, MCP gateway and Minion agent handling about 15% of production PRs

Uber's Uday said every model request goes through one gateway exposing OpenAI- and Anthropic-compatible APIs, and its MCP gateway, started around April 2025, auto-generates MCP servers from service specs.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: The Pragmatic Engineer: Uber’s AI infra, from the inside

Google paused open-source vulnerability rewards on Oct. 1; PageBreak agent reportedly found about 500 vulnerabilities

An IBM Technology host said Google paused its open-source vulnerability rewards program on Oct. 1 after AI-generated reports full of hallucinations, with the open-source part frozen until Q1 2027.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: IBM Technology: AI slop stalls bug bounties, malware hides in IoT and agents hit…

Models and learning

Claude Motion beta is limited to Team and Enterprise plans; presenter's one-shot timeline took about four minutes

The AI Advantage presenter said Claude Motion, under the Artifacts sidebar with a Dashboard tab, is included only in Team and Enterprise accounts for now.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: The AI Advantage: New Claude Motion is Insane (high hype)

Apple Mac Studio with M5 Ultra begins shipping; relayed estimates put RTX 5090 ahead on Qwen 3.8 27B

A Hugging Face speaker said the Mac Studio with M5 Ultra started shipping with up to 256 GB of memory, 512 GB coming, 1.2 TB/s bandwidth and about 110 TFLOPS.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Hugging Face: RTX 5090 vs M5 Ultra: Which AI Powerhouse Wins?

Speakers cite harness-only changes, such as Grok Code Fast 1 rising from 6.7% to almost 70%, as evidence harness matters

Leon van Zyl relayed a post saying Grok Code Fast 1 went from 6.7% to almost 70% with tool changes only, and a LangChain experiment in which GPT-5.2 Codex rose from about 53% to about 66% with its deep agent CLI.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: Leon van Zyl: Same Model, 10x the Score: The Hidden AI Bottleneck

Mastra engineer says team reviews plans more than PRs and uses a second model to review

A Mastra speaker said his team reviews implementation plans more than pull requests, builds with Claude and reviews with Codex or the reverse until the reviewer has no comments, runs visual QA agents per PR with cached steps to limit cost, and uses background agents that watch production signals and open fix tasks.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Mastra: LLMs Can't Write Great Software, So We Verify Everything - Ohans…

REA MCP server and other AI-built reverse-engineering repos draw expected legal-pushback comments

Fahd Mirza described REA, a local MCP server using Ghidra that gathers evidence from binaries, Electron apps and websites for coding agents.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Fahd Mirza: REA: Reverse Engineer Anything Tutorial with Hands-on Demo

Microsoft demo of Azure Container Apps sandboxes: about one-second start, default-deny egress, 8,461 ms cold versus 2,842 ms primed

A Microsoft Reactor presenter said each Azure Container Apps sandbox has its own kernel and file system, starts in about a second and has egress denied by default.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Microsoft Reactor: Creating Custom Code Execution Engine Sandboxes For Your Loops And…

Microsoft team cites 60% lead-time cut with AI SDLC; Copilot modernization demo has one of three PRs fail

A Microsoft Reactor speaker cited one internal team with a 60% reduction in lead time and 38% fewer pre-production bugs per feature after adopting an AI software development lifecycle.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Microsoft Reactor: The Agentic SDLC: Making Application Modernization Continuous

Microsoft Research presents fixed-point score matching for training samplers without data

A Microsoft Research speaker presented a method that trains diffusion samplers by regressing on target scores of the model's own samples held in a replay buffer.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Microsoft Research: Scalable Training of Diffusion Samplers via Fixed-point Score Matching

Coverage

Window: videos uploaded in the 24 hours to Oct. 15, 02:13 UTC. Of 448 videos inventoried from monitored channels over the four-day scan, 97 were read for this edition, all uploaded inside the window, so there are no late finds today; 283 had been covered in earlier editions, 6 had no transcript available, 15 were live or premiere videos deferred and 62 were under 45 seconds. The collector saved 97 of 103 queued transcripts and completed normally.

Sources reviewed (97 videos)

All 97 videos read for this edition, grouped by channel. Tags: what the video was, plus sponsored (paid placement), vendor (a company on its own product), self-promo (the creator's own offer), high hype or partial transcript where they apply.