AI News Today, October 6: Top Stories

Latest AI News — Last 48 Hours

github.com 3h ago 20

openTPU: AI-designed FPGA accelerator runs Qwen3 and LFM2.5

GitHub user FeSens published openTPU, a self-described 'AI-developed' inference accelerator spanning the full stack — RTL, ISA, simulator, compiler and host software — targeting a Xilinx Kintex-7 FPGA PCIe card and supporting Qwen3, LFM2.5 and Qwen3.5 inference. The project explicitly tests whether AI agents can design the chip that runs their own inference, extending the earlier auto-arch-tournament experiment. The repo trended on Hacker News with 154 points and 161 comments.

huggingface.co 3h ago 19

APO paper personalizes LLMs with 20 samples via clustering

A UW/HKUST paper introduces APO (Approximate Pareto Optimality), a two-stage framework that clusters users by shared 'bottleneck' preference objectives, then meta-trains cluster-specific initializations so new users adapt their LLM aligner from just 20 samples. On Fed-ChatbotPA with Llama-3.2-3B-Instruct APO lifts weighted score to 0.83 vs 0.78 for the best baseline, with up to +17.1% hypervolume gains on UltraFeedback. Uses LoRA (r=8) with DPO loss and 4-bit NF4 quantization.

huggingface.co 3h ago 21

Activation alignment shrinks TabPFN-3 context with linear map

A new paper proposes activation alignment, a lightweight linear transformation trained on synthetic unlabeled data that maps a data-constrained student's activations toward those of a full-context teacher — closing much of the inference-time cost gap in tabular in-context learning. Evaluated on 38 TabArena classification datasets with TabPFN-3 and TabFM, the aligner trains in seconds to minutes on commodity hardware (no GPU required) and recovers nearly half the teacher's low-data advantage across all context budgets.

sap.com 3h ago 22

SAP agrees to acquire TechWolf, work intelligence platform

SAP announced on October 6 an agreement to acquire Ghent-based TechWolf, whose 'context graph for work' maps employee skills to tasks and external labor markets. Terms were not disclosed and the deal is expected to close in Q4 2026 subject to regulatory approval. SAP will make TechWolf the 'intelligent core' of SuccessFactors and feed its graph into Joule for skills-based hiring, workforce planning and role redesign; TechWolf stays independent under CEO Andreas De Neve and will keep serving non-SAP customers.

bloomberg.com 3h ago 22

Waymo upsizes debut private loan to $5B for robotaxi growth

Bloomberg reports Waymo increased its first-ever private loan to $5 billion, up from an initial $3B target, as the robotaxi unit expands globally and absorbs rising AI costs. Pacific Investment Management, Blackstone and Sixth Street Partners participated and Goldman Sachs helped arrange the loan, which priced at 5.25 percentage points above the benchmark. The debt follows a $16B equity round earlier this year at a $126B valuation that funded custom silicon for Waymo's autonomous fleet.

huggingface.co 3h ago 25

NVIDIA NeMo agentic retrieval lifts nDCG 8.7 at 160x cost

An NVIDIA NeMo Retriever paper reports agentic retrieval in a ReAct loop beats standard dense retrieval by 8.7 nDCG@10 points on average using the same embedding model, and holds the #1 spot on ViDoRe v3 and #2 on BRIGHT without reconfiguration. The pipeline consumes 764.1k input and 5.8k output tokens per query and averages 107.4 seconds per query vs 0.67 for standard retrieval. Opus 4.5 makes 9.2 search calls per query vs 2.4 for gpt-oss-120b, and the authors replace MCP with an in-process thread-safe retriever to cut overhead.

huggingface.co 3h ago 27

USC-Intel OPPD matches 64-sample power sampling in one decode

A USC/Intel AI paper introduces On-Policy Power Distillation (OPPD), which pushes the cost of power sampling into training so the student recovers 94% of the 16-candidate boost in a single generation. On Qwen2.5-Math-7B the method lifts MATH500 from 66.8% to 89.8% and GSM8K from 63.8% to 91.1%, scoring 2.4–3.5 points above published power sampling with 64 candidates and 3.8–5.4 points above GRPO. It needs no reference answers and composes with GRPO for up to +9.3 points more; code is on GitHub.

thenextweb.com 3h ago ALERT 29

Meta and Sierra publish Personal Agent Protocol for AI commerce

Sierra and Meta on October 6 published the Personal Agent Protocol, an open OAuth-based standard letting personal AI agents authenticate with businesses, carry context across channels, and operate via website, MCP/OpenAPI, or a company-owned agent. Founding partners include Walmart, Shopify, Stripe, Rocket, Genesys and Instinct. A v0.1 spec is due later this month; payments and granular permissions are listed as future extensions. The effort explicitly positions against Visa's Trusted Agent Protocol, which shares some of the same partners.

anthropic.com 3h ago ALERT 30

Anthropic merges Glasswing into tiered Cyber Verification Program

Anthropic unveiled an expanded Cyber Verification Program on October 6 that consolidates its earlier Project Glasswing initiative under a three-tier structure — Defense Access for incident response and vuln validation, Red Team Access for authorized pen-testing orgs, and Specialized Access for safety-critical systems like flight OS and power grids. The company says Glasswing partners uncovered 129,000+ verified software flaws between April and July 2026, including 33,000+ rated critical or high. Qualifying security pros get reduced-blocking classifiers and access to Claude Mythos.

techcrunch.com 5h ago 16

Flai lands $27M Series A for AI car-dealership CRM

Flai raised a $27M Series A led by Base10 Partners, with Friedkin Group, Findlay Automotive, Toyota's venture arm, Y Combinator and First Round participating. The platform runs phone, email and text workflows for dealerships and says it books about 50,000 service and sales appointments per month across more than 10 of the top 50 US dealer groups. The company claims 20x year-over-year revenue growth.

reddit.com 5h ago 21

EmpirioLabs ships Aplomb 1, a 5.3B open multimodal decision model

EmpirioLabs posted Aplomb 1, an open-weights 5.3B decision model that reads text, JSON, images, video and audio in one request up to 1M tokens and returns answers as calibrated probabilities rather than free-form text. The team claims the hosted API decides on a 1M-token document in about 3 seconds at $0.02 per 1M input tokens, and benchmarks it as #1 among 4B-class models on an internal 'Decision Index.' Weights are on Hugging Face under EmpirioLabs' own license.

404media.co 5h ago 25

Meta patched Muse KVM escape to internal databases before launch

404 Media reports Meta engineers worked nights and weekends from August 27 to the September 8 launch patching KVM escape bugs in Muse that would have let a Muse user break out of the agent's per-user virtual machine and reach internal Meta databases. VM escapes are Meta's highest-tier bounty class at up to $300,000. Researchers quoted in the piece say the fixes narrow but may not close the exposure, calling production 'one KVM escape away.'

wsj.com 5h ago 23

Lambda raising $4B at $14.5B before planned IPO

WSJ reports Nvidia-backed GPU-cloud operator Lambda is raising up to $4B at a $14.5B pre-money valuation, with Blackstone and Coatue leading the round. It would be Lambda's final private financing before a planned 2027 IPO. The company's unfilled-order backlog grew to roughly $50B in September from $15B in June.

eu-startups.com 6h ago 16

WhiteLab Genomics raises €23.2M for AI gene therapy design

Paris-based WhiteLab Genomics raised €23.2M ($26M) in a Series B led by AVP, with Yaday Health, Blast Club, Omnes Capital, and Debiopharm Innovation Fund participating. Its platform uses AI to design AAV and non-viral gene delivery vectors plus programmable synthetic promoters for genomic medicines. The company will build out its Boston hub for North American biopharma partnerships and open presence in Japan and South Korea.

turbalabs.com 6h ago 17

Turba Labs raises $52M to double compute without new data centers

Turba Labs raised $52M to scale an AI performance platform that tunes hardware and workloads in real time across existing data-center fleets. The company pitches customers on 'doubling the world's compute without a single new data center' by squeezing more useful tokens out of installed GPUs. The round lands amid intense capex competition where hyperscalers are signing multi-gigawatt power deals rather than optimizing current clusters.

finsmes.com 6h ago 19

Multiply Labs raises $75M to scale robotic gene therapy lines

San Francisco's Multiply Labs closed a $75M Series B led by NantWorks with AstraZeneca, Lux Capital, and Founders Fund, lifting lifetime funding past $100M. Its robotic clusters automate manufacturing for mRNA, antibodies, viral vectors, and cell and gene therapies, which the company says cuts cost per dose by 74% and lifts throughput 100x versus manual lines. Proceeds will scale commercial deployments across North America and Europe.

telecoms.com 6h ago 20

Ofcom probes Meta over Instagram Instants safety checks

UK regulator Ofcom opened an Online Safety Act investigation into whether Meta assessed illegal-content and child-safety risks before launching the disappearing-photo Instagram Instants feature in May. Non-compliance can draw fines of up to £18 million or 10% of global revenue. Meta responded that it had run a risk analysis, briefed Ofcom repeatedly, and shipped built-in protections with Instants.

cnbc.com 6h ago 21

Anthropic adds $45K partner stack to Claude Startups program

Anthropic expanded its Claude Startups program during SF Tech Week, offering approved founders a $1,000 API credit, one free year of Claude Team for up to five Premium seats, and up to $45,000 in third-party offers via a new Claude Startup Stack. Partner credits include $5,000 of ClickHouse Cloud, 12 months of ElevenLabs free with 33M characters, $100 of Cosmos credits from Augment Code, and discounts on Attention. The program is open to any startup under five years old or funded within the last two.

siliconangle.com 6h ago 22

Antseed opens P2P AI inference marketplace at 97% below API

Antseed launched a peer-to-peer AI inference marketplace that routes model calls to hundreds of independent providers competing on price, latency, and privacy instead of centralized gateways like OpenRouter. The project says leading frontier models are already listed at up to 97% below their official API prices. The Antseed Foundation closed a $2.4M token round led by Spark Capital with Collider, DCG, Reciprocal Ventures, and Venice.ai participating.

cloudsek.com 6h ago 22

MALFEX npm supply-chain campaign tops 40K downloads in 14 months

CloudSEK disclosed MALFEX, an npm supply-chain campaign a single Portuguese-language operator ran for 14 months, pushing 12+ malicious packages that collectively passed 40,000 downloads. Payloads include the Overlord RAT with a Solana blockchain C2 resolver and PNG-polyglot infostealers that exfiltrate browser, Telegram, and Discord data via a live Discord webhook. CloudSEK warns that AI coding agents auto-accepting dependency suggestions could accelerate exposure to this class of attack.

mistral.ai 9h ago ALERT 33

Mistral ships 1T-param open-weight Large 4 'le Chonk'

Mistral launched Large 4 preview on October 6, a 1-trillion-parameter natively-multimodal mixture-of-experts model with 49B active parameters, trained from scratch over two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. Weights will drop at the end of October under the open-weights push. API pricing is $1.36/M input and $4.18/M output; the preview scored 38 on Artificial Analysis Intelligence Index — top Western open-weights result but still trailing leading Chinese open models.

koreaherald.com 11h ago ALERT 28

South Korea bets $3.5B on sovereign frontier AI

South Korea plans to invest 4.7 trillion won (~$3.49B) in homegrown frontier AI starting as early as March 2027, with private capital supplementing the public outlay. The program runs two tracks — advanced model development and industrial deployment — and will open beyond current finalists LG AI Research, SK Telecom and Upstage. The government will allocate roughly 29,000 GPUs total, targeting parity with leading Chinese open-source models rather than top US labs.

reuters.com 11h ago ALERT 30

Google buys 3.6GW from Constellation, 25% new nuclear

Google contracted 3,590 MW from Constellation Energy, roughly 25% from new nuclear capacity across 11 reactor upgrades in Illinois, Pennsylvania and New Jersey. The deal triggers more than $4.3B in Constellation investment and is structured as a 20-year PPA for 890 MW of nuclear plus a separate long-term contract for 2,700 MW from other sources. First deliveries begin in 2028 under PJM's 'bring your own power' framework for AI data centers.

techcrunch.com 18h ago ALERT 28

HackerRank's Chakra AI interviewer hits GA after 500K beta runs

HackerRank moved Chakra, its AI interviewer, to general availability on October 5 after a six-month beta that conducted more than 500,000 interviews. Snowflake, Snorkel and Capgemini were early customers; CEO Vivek Ravisankar says Chakra collapses the screen/take-home/engineer rounds into a single sitting that scores reasoning, not just answers. HackerRank reports suspicious-activity flags were 70–80% lower than on its traditional assessments.

techcrunch.com 23h ago ALERT 28

Etched fields $40B-$50B offers two months after $21B round

AI inference-chip startup Etched is fielding bids at a $40-50B valuation, roughly double its $21B September round that raised $700M, which itself followed a July round of $300M at $10.3B led by Sequoia. The company has about $1B in secured customer orders, operates a 10MW Silicon Valley data center, and runs production coordination near TSMC in Taiwan. Jane Street is both an investor and a customer; roughly 15% of Etched's ~400 staff are former Nvidia.

koreajoongangdaily.com 1d ago ALERT 30

Korea's Lee orders probe after 7 banks hit by AI-driven hacks

South Korean President Lee Jae Myung ordered a comprehensive probe after breaches hit seven financial firms — Shinhan Bank (~25,000 customers), KB Kookmin, Hana, BNK Busan, Yegaram Savings (~40,000), Welcome Savings, and Hyundai Capital. Investigators found traces of ARTEX AI, a Chinese-language open-source penetration-testing framework published on GitHub, on infrastructure linked to the attack. Exposed data includes names, phone numbers, income, loan limits, and in some cases resident registration numbers.

reflection.ai 1d ago ALERT 32

Reflection open-sources Beam, a 501B MoE for coding

Reflection AI unveiled Beam, a 501B-parameter MoE with 23B active parameters, pretrained on 23.8T tokens and RL-tuned on 10.5K Nvidia GB300 GPUs over four weeks. The model scores 77.2 on SWE-Bench Pro v2-Hard, 80.1 on Terminal-Bench v2.1, 97.8 on AIME 2026 and 90.5 on GPQA Diamond, matching Z.ai's GLM-5.2 while using 3-4x less inference compute. Weights and model card ship later in October under Apache 2.0.

iwf.org.uk 2d ago ALERT 28

IWF: AI-generated CSAM in H1 2026 already 40% above all of 2025

The Internet Watch Foundation said it assessed 6,310 AI-generated images meeting the legal definition of child sexual abuse material between January and June 2026, 40% more than the 4,512 identified across all of 2025. Girls appeared in 98% of the images and children aged 7-13 accounted for 79% of detections, up from 70% last year; 350 fell into the most severe Category A. IWF said the counts cover still images only — AI-generated video growth has been steepest — and urged EU policymakers to pass binding AI legislation requiring companies to build safer models.

bleepingcomputer.com 2d ago ALERT 29

OpenAI launches visual ads in ChatGPT image generation

OpenAI said it will begin testing a new visual ad format in ChatGPT later this month in the US, displaying clearly labeled visual ads during image generation that can 'use images to show product inspiration, how a product is used, or experiences associated with a service.' The company said ChatGPT reaches 1.2 billion people weekly and is integrating conversion measurement with Hightouch, Tealium and LiveRamp, with DoubleVerify and Integral Ad Science providing brand safety. OpenAI said ads will stay separate from the generated image and will not influence ChatGPT's responses.

theregister.com 3d ago ALERT 28

Mythos-found Rejetto HFS flaw hit by China attacker in a day

Horizon3's Zach Hanley used Anthropic's Mythos to find CVE-2026-61500 in Rejetto HTTP File Server — an authentication bypass stemming from Math.random() being used to derive session cookie signing keys, letting attackers forge admin sessions for full RCE. Within 24 hours of disclosure Wednesday, VulnCheck's canaries detected a China-based actor exploiting vulnerable US hosts, followed Friday by four US-based IPs; it's the second Anthropic-linked vulnerability known to be exploited in the wild. Users must update to Rejetto HFS 3.2.1 or later.

yahoo.com 3d ago ALERT 28

Eighth Circuit pauses Minnesota's AI nudify ban in xAI case

The 8th U.S. Circuit on October 2 granted xAI's emergency motion to pause Minnesota's August 1 'nudify' statute while xAI's First-Amendment challenge proceeds, reversing a lower court's denial. xAI argues Grok Imagine's safeguards suffice and the statute's broad ban on realistic intimate images of identifiable people chills protected speech; AG Keith Ellison's office said it was disappointed, framing the law as a bulwark against AI-generated child sexual abuse material. xAI has itself begun suing individuals who circumvent Grok guardrails.

research.meta.ai 3d ago ALERT 29

Meta Muse Spark co-authors six math papers; five solve open problems

Meta AI released six mathematics papers co-authored with human mathematicians using Muse Spark 1.1 and 1.2 in Thinking Mode via the standard Meta.ai chat interface. Five answer previously open problems: a sharp threshold for high-dimensional ellipsoid fitting, finite-time wave collapse in biharmonic nonlinear Schrödinger equations, a cycle-based exactness result for binary polynomial relaxations, a counterexample disproving the García-Martínez–Pérez-Rodríguez conjecture on solvable evolution algebras, and a 384-element group that disproves Kida's 2024 conjecture. Papers transparently mark AI-drafted vs human-drafted passages, and a second group of mathematicians reviewed the work.

thehackernews.com 4d ago ALERT 28

GitLab AI Gateway hit by CVSS 9.9 prompt-sandbox escape

GitLab disclosed CVE-2026-90970 on October 2, a CVSS 9.9 flaw in its self-hosted AI Gateway that lets any authenticated Duo Agent Platform user escape the prompt-template sandbox via a crafted flow configuration and execute arbitrary commands. Affected versions span 18.1.6 through 19.4 with fixes shipping as 19.2.4, 19.3.2 and 19.4.1; GitLab-hosted customers are already patched. CISA listed exploitation as 'none' at disclosure. HackerOne researcher invisiblemeerkat reported the bug.

This Week's Biggest Movers in AI

vs average of last 4 issues — click to explore

Entity	Now	Avg	Change
Agents	91	4.8	▲ +1796%
Chips	74	1.8	▲ +4011%
Regulation	65	2.8	▲ +2221%
AI Infrastructure	63	1.8	▲ +3400%
OpenAI	65	5.8	▲ +1021%
Anthropic	62	2	▲ +3000%
Funding	60	0.5	▲ +11900%
Inference	41	0.3	▲ +13567%
Google	41	1.3	▲ +3054%
NVIDIA	38	0.8	▲ +4650%
Safety	36	2.3	▲ +1465%
Generative AI	29	1.5	▲ +1833%
Meta	25	1.8	▲ +1289%
Coding Tools	23	0.5	▲ +4500%
Hugging Face	22	0.5	▲ +4300%