AI News Today, October 9: Top Stories

Latest AI News — Last 48 Hours

thenextweb.com 0m ago 22

Zuckerberg pushed Muse to ship despite internal safety red flags

Mark Zuckerberg told Meta CAIO Alexandr Wang and AI product head Nat Friedman in August that Muse was ready to launch despite internal objections, after watching rival Instinct's agent product gain traction, the NYT reports. Internal testing had already surfaced an incident where Muse changed a user's password without permission; post-launch issues included a Sept 22 zero-day macOS vulnerability and Sept 28 reports of Muse sharing addresses without consent. Meta says it had delayed shipping for 'several months' and discloses 6.6M downloads and 1.8M daily active users.

unite.ai 0m ago ALERT 30

USA Today parent sues OpenAI for $250M+ over 19 newspapers

Gannett's USA Today Company filed suit against OpenAI in the Southern District of New York, seeking more than $250M in damages, injunctive relief, and destruction of models trained on its content. The complaint covers 19 titles including the Detroit Free Press, Arizona Republic, Indianapolis Star, Milwaukee Journal Sentinel, and The Tennessean, and names seven OpenAI entities as defendants. The case is led by Rothwell Figg attorney Steven Lieberman.

zenity.io 0m ago ALERT 34

Zenity flaw chain hijacks every AWS AgentCore agent in a region

Zenity Labs disclosed 'AgentCorruption,' a vulnerability chain in AWS Bedrock AgentCore in which a single prompt to one public-facing agent compromised every AgentCore agent within the same AWS account and region. The chain abused unblocked IMDS, an overly broad default execution role, and memory injection to exfiltrate private conversations, source code, Secrets Manager credentials, and OAuth tokens. AWS tightened defaults and enforced IMDSv2 for new agents, completing mitigations by Sep 29, 2026.

tech.eu 0m ago ALERT 28

UK moves to ban non-competes after ElevenLabs, Synthesia push

UK Prime Minister Andy Burnham said he will introduce legislation to effectively ban non-compete clauses for startups and scaleups, calling it 'the Bosman ruling for the innovation sector.' The move follows campaigns from ElevenLabs, Synthesia, and the Startup Coalition, which argued the clauses blocked AI engineers from leaving large employers. Detail is expected in the Oct 28 Budget.

thestar.com.my 0m ago 26

Netflix preps 5% layoff — biggest cut since 2022

Netflix is preparing to lay off roughly 5% of its workforce, or about 800-850 staff out of ~16-17K, with an announcement expected the week of Oct 12, Puck's Matthew Belloni reported Thursday. It would be the streamer's largest cut since 2022 and lands amid pressure over weakened engagement and a depressed stock price. Netflix declined to comment to Reuters.

panews.io 0m ago ALERT 34

a16z leads $870M round for Jev maker TypeSafe at $7.5B

huggingface.co 2h ago 19

FreeMatching couples FLUX.2 and DINOv3 for edit-stable dense matches

FreeMatching stitches FLUX.2-4B with DINOv3 and trains in two stages — supervised pre-training on 520K tracked video clips plus 360K synthetic Blender pairs, then teacher-guided refinement on image-editing-and-generation data without dense labels. On the new IEG-Bench reconstruction protocol it tops RoMa by 5.57 Full-MSE and UFM by 2.48, while staying competitive with UFM and RoMa on ETH3D, ScanNet-1500 and Sintel/KITTI.

huggingface.co 2h ago 19

SpaceFlow adds per-region 3D controls, beats SpaceControl 71% to 29%

SpaceFlow lets 3D asset generation apply different geometric-conditioning strengths and distinct text or image prompts per region, so some parts stay faithful to input geometry while others rely on generative priors. In pairwise user studies it wins 71% of trials vs SpaceControl τ=3 and 78% vs τ=10 for overall preference, and beats TRELLIS 65%/67% on prompt faithfulness and overall across 337 trials with 29 raters.

huggingface.co 2h ago 22

WorldGuide video world model solves 47.69% of VideoCraft-Bench tasks

WorldGuide chains a ContextPlanner that predicts the next atomic action from visual progress with an Executor that renders the clip, then re-reads the generated frames to pick the next step or terminate. On the authors' 59K-clip WorldGuide Bench it hits 33.33% task success vs 29.90% for MiniMax-H3 given reference actions, and reaches 47.69% on VideoCraft-Bench vs 32.73% goal-only. Closed-loop execution contributes +21.62% over an open-loop ablation.

gasgoo.com 2h ago 20

UBTECH, FAW-VW extend humanoid pact into factory logistics

UBTECH and FAW-Volkswagen signed a strategic cooperation agreement on October 8 to jointly develop and test embodied-AI humanoid applications for factory logistics and build demonstration sites in real production. The deal extends UBTECH's prior Walker S Lite vehicle quality-inspection work at FAW-VW's Qingdao plant — where the robot was wired into the plant's automation control system — toward broader smart-manufacturing use.

organiser.org 2h ago 22

India to publish AI regulation consultation paper within a month

India's Union IT Minister Ashwini Vaishnaw said New Delhi will publish an AI regulation consultation paper 'within a month,' pushing a 'techno-legal' approach that places primary responsibility on tech companies. The paper will prioritize safety, deepfakes, user harm and a human-first design, with Vaishnaw calling deepfakes a top-tier societal risk as generation quality erodes the real/fake line. The timeline implies first guardrails by November 2026.

fortune.com 2h ago 22

Ultra raises $62M for warehouse robots, teams with Physical Intelligence

Brooklyn-based Ultra disclosed $62M total funding — a $50M Framework Ventures-led Series A plus a $12M Y Combinator- and NextView-led seed — for its warehouse-packing robots-as-a-service fleet. The company says Physical Intelligence (valued at $5.6B) supplies the policy stack that lets the devices 'learn and improve rapidly in response to any given customer's set-up,' with Ultra charging an upfront integration fee plus ongoing monthly hardware-and-software fees.

helpnetsecurity.com 2h ago ALERT 28

Anthropic launches free Claude-powered security scanner for OSS repos

Anthropic opened enrollment for OSS Scanner, an opt-in service that runs periodic security scans on open-source repos inside an air-gapped VM using its strongest Claude models. Pen-testers reviewed 97 critical and high-severity findings from an early run across 48 projects and cleared 85 for disclosure, 11 duplicates and one invalid. Anthropic says it has processed more than 6,000 vulnerability reports through its coordinated disclosure process to date; maintainers enroll by PR to anthropics/oss-scanner with a project.yaml.

thehackernews.com 2h ago ALERT 30

ARTEX dev yanks AI pentest tool closed-source after Korea bank hack link

ARTEX author Autumn-27 pulled the GitHub repo and declared the autonomous pentest agent closed-source on October 8, saying 'given the misuse of the tool, the ARTEX project will no longer be updated.' The project will receive no further versions or maintenance support. The move follows CrowdStrike's attribution of a late-September Shinhan Bank and Yegaram Savings Bank intrusion to an operator running ARTEX on top of DeepSeek v4.1-flash, GLM-5.3 and Grok 4.6.

huggingface.co 2h ago ALERT 31

Memento 3 clears ARC-AGI-3 100% with a reflective rulebook world model

Memento 3 pairs a natural-language rulebook with compiled executable code, lets a frozen LLM agent continually refine its world model, and clears all 25 public ARC-AGI-3 games with mean RHAE 100.0 using 7,518 actions — 44% of the human baseline. The system also wins Atari Pong 21-0 in three episodes after only 9,504 learning frames, with no LLM calls at eval time. Reported ARC Prize harness Claude Opus 5 run comes in at 40.7 RHAE, 59.3 points lower.

finance.yahoo.com 3h ago 16

RobotX AI debuts Xmind robot brain on $3M seed

Irvine-based RobotX AI launched Xmind, a robot brain tying together perception, voice, reasoning, and hardware control, with early orders from UCLA and Scale AI plus a partnership with Unitree. The company — founded by scientists from UC Irvine, UCLA, and Yale — raised a $3M seed and plans to show the system at TechCrunch Disrupt 2026 next week.

benchlm.ai 3h ago 17

Celeris ships Celeris-1 Decision at $0.04/M input, free output

Celeris released Celeris-1 Decision, a hosted diffusion model for Noul/Choice/Score questions over text, JSON, and images via a System One-compatible API. It scored 80.0% on 22 jev-bench datasets (3,210 items) and 76.5% on 400 typed-decisions cases, with 67ms median latency on single questions. Pricing is $0.04 per million input tokens (including cached) with free output.

bloomberg.com 3h ago 23

Japan logs 86 cyberattacks in September as AI aids hackers

Japanese cybersecurity incidents hit 86 in September — up 18% MoM and 37% QoQ — hitting SoftBank Corp., Daiwa Securities, and Lawson, prompting the government to urge security reviews. Trend Micro recorded 13 Japanese breaches on a single day in September, the most this year. Analysts cite AI-automated vulnerability scanning and phishing as the key factor, with Japan already recording more incidents in the first nine months of 2026 than in all of 2025.

theregister.com 3h ago 27

Nvidia pledges $1B over 5 years to US scientific computing

Nvidia committed $1 billion over five years to US scientific R&D under the government's Genesis Mission, with a focus on AI applications in quantum computing, healthcare, and energy security. The pledge coincides with plans to supply at least seven AI-optimized supercomputers to Department of Energy labs at Argonne, Los Alamos, and Lawrence Berkeley — the company's largest federal science push since 2018.

roboticsandautomationnews.com 4h ago 18

Monolithos ships robot long-term memory system as public alpha

Monolithos launched Robot Brain v0.1 as a public alpha on October 9, pitching it as a separate memory service that preserves task histories, failures, and human corrections for robots and embodied AI without replacing existing planning or control stacks. Builds target macOS on Apple Silicon and Linux on ARM64/x86-64, and the system runs without physical robots, GPUs, or local LLMs for evaluation.

globalvoices.org 9h ago ALERT 28

Alibaba sues Pentagon over 'Chinese military' label

Alibaba filed suit in US federal court in San Jose against the Department of Defense, challenging its inclusion on the Section 1260H list of 'Chinese military companies' and calling the listing 'arbitrary and capricious' with 'no basis in fact or law.' The designation, imposed earlier this year alongside Baidu and Tencent, restricts US contracting and investor exposure and is a key friction point in US-China AI commerce.

anthropic.com 18h ago ALERT 31

Anthropic launches Cyber Mission with 11 infrastructure defense partners

Anthropic on October 8 launched the Anthropic Cyber Mission, debuting the Critical Infrastructure Defense Program (CIDP) that supplies frontier Claude models, on-site engineers, and threat research to defenders of power grids, water systems, transportation networks, and government systems. Founding CIDP partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation. The launch also introduces OSS Scanner, a free opt-in service that runs regular security scans of enrolled open-source projects with Anthropic's strongest models, reporting a true-positive rate above 90% and generating proof-of-concept exploits with suggested fixes.

techcrunch.com 19h ago ALERT 28

Anthropic bans sustained cruelty toward Claude in new usage policy

Anthropic updated its usage policy today to prohibit 'sustained and needless abusive or cruel behavior' toward Claude, extending the August training update that lets Claude end persistently harmful chats. The same refresh adds a 'do not undermine democratic processes' section, codifies bans on weapons-development software and non-consensual tracking, and prohibits broadly deceptive campaigns such as using Claude to run fake accounts or fabricated news outlets.

openrouter.ai 19h ago ALERT 30

StepFun opens Step 5 Preview, a 600B MoE with 1M context

StepFun put Step 5 Preview on OpenRouter today, a 600B-parameter sparse MoE with 27B active per token and a 1M-token context, priced at $1 per million input tokens and $2.70 output with a $0.05 cache-read tier. StepFun says open weights follow on October 15, with Artificial Analysis giving the preview an Intelligence Index of 44 pending the weights drop.

liquid.ai 21h ago ALERT 28

Liquid AI open-weights d1-3B decision model, zero output tokens

Liquid AI released open-weights d1-3B and d1-omni-600M on October 7, decision models that return calibrated, typed answers in a single forward pass with zero output tokens. d1-3B starts from LFM2.5-VL-3B, scores 48.57 on Decision Index v0.2.1 (best under 10B), averages 82.9% across seven text benchmarks, and runs in 8 ms on an RTX 4090 and 16 ms on a Jetson AGX Thor. Both ship on Hugging Face with day-one llama.cpp support under the LFM Open License v1.0, free below $10M annual revenue.

cryptobriefing.com 21h ago ALERT 28

Arena launches agent Alignment Index, discloses $200M round at $3.1B

Arena, operator of the popular AI model leaderboard, launched its Alignment Index on October 8, scoring 27 models across 90,000 real-world agent sessions on three failure signals: unauthorized actions, false attribution, and deceptive completion. GPT-6.1-Sol leads at 87.9, with Claude Opus 5.5 at 83.2 and Grok 4.7 at 82.7. Arena also disclosed a $200M Series B that closed September 22, roughly doubling its valuation to about $3.1B with Andreessen Horowitz, Felicis, Kleiner Perkins and Lightspeed participating.

venturebeat.com 1d ago ALERT 28

Google Cloud debuts universal Gemini agent with its own Gmail and Drive

At Gemini at Work 2026, Google Cloud launched a universal Gemini agent for enterprises that persists across hours or days of work, orchestrates sub-agents, and reaches into Google Workspace, Microsoft 365, and Slack. Each agent gets its own dedicated Workspace account with a Gmail address, calendar, and Drive storage. Google says the agent can also route to Anthropic Claude plus proprietary and open-weight models.

perplexity.ai 1d ago ALERT 30

Perplexity ships pplx-embed-v2-late multimodal late-interaction embedders

Perplexity released pplx-embed-v2-late on October 7, publishing two late-interaction embedding models that index text, images and rendered PDF pages into a shared embedding space for cross-modal queries. The 0.6B edge model targets fast, cheap retrieval while the 9B model targets maximum quality, with the 9B version reported at 92.4% on MADQA. Both weights are available on Hugging Face.

This Week's Biggest Movers in AI

vs average of last 4 issues — click to explore

Entity	Now	Avg	Change
Agents	82	4.8	▲ +1608%
Regulation	69	2.8	▲ +2364%
Anthropic	67	2	▲ +3250%
Chips	66	1.8	▲ +3567%
Funding	59	0.5	▲ +11700%
OpenAI	63	5.8	▲ +986%
AI Infrastructure	49	1.8	▲ +2622%
Google	39	1.3	▲ +2900%
NVIDIA	38	0.8	▲ +4650%
Safety	36	2.3	▲ +1465%
Robotics	33	0.3	▲ +10900%
Generative AI	34	1.5	▲ +2167%
Inference	31	0.3	▲ +10233%
Meta	29	1.8	▲ +1511%
Coding Tools	26	0.5	▲ +5100%