AI News Today, October 30: Top Stories

Latest AI News — Last 48 Hours

pymnts.com 1m ago 19

UBS employees now run 12,700 personal AI agents, use cases up 81%

UBS said in its Q3 results on Oct 29 that employees now run more than 12,700 personal AI agents in their daily work, with 608 live AI use cases in production (up 81% YoY) and 1,090+ more in development. The bank's AI tooling reaches 18,000 software developers; its STAAT financial-advisor tool has generated 30 million advisory insights. The scale is pushing agent governance and audit-trail inventory onto the board agenda.

bloomberg.com 1m ago 23

Amazon lines up $80B secured supplier loan to fund AI hardware

Amazon disclosed in its Oct 30 quarterly filing a secured loan facility of up to ~$80B against technology-infrastructure assets, with a drawdown window from Q1 2027 through Q3 2028. The facility is designed to help fund Amazon's AI capex, which stays at ~$220B for 2026 and is expected to rise meaningfully in 2027; trailing-12-month free cash flow sits at negative $42.4B. AWS CEO Matt Garman framed it as 'responsibly funding' AI investments through creative financing Amazon had previously avoided.

huggingface.co 1m ago 22

CMU and Aurora paper collapses all foundation model training into one algorithm

Song et al. (CMU, Johns Hopkins, Aurora Innovation) propose unified training (arXiv 2610.40564, Oct 30) that treats pretraining, midtraining, SFT and RL as evolving mixtures of a single maximum-likelihood objective with per-token weights. RL updates become likelihood fits to success-conditioned distributions via MaxRL, letting one algorithm span the whole lifecycle. Experiments on 124M logic-puzzle and 0.76B math models show higher accuracy, coverage and out-of-distribution generalization than staged pipelines at the same compute budget.

arxiv.org 1m ago 27

LeCun paper proves Platonic Hypothesis for JEPAs, 42x sample efficiency

Randall Balestriero and Yann LeCun's new paper (arXiv 2610.39302, Oct 29) proves that JEPAs trained separately on different modalities converge to the same representation up to one rotation when positive pairs share world states — turning the Platonic Representation Hypothesis into a formal theorem. A 109M-param text JEPA beats language models up to 72B on the PRH benchmark, needing 42x fewer samples than next-token prediction to reach equivalent cross-modal alignment. The paper reframes LLM pretraining as a 'bad-view JEPA,' explaining why language models align poorly with vision despite scale.

ft.com 1m ago 22

Blackstone says Anthropic likely lands junk rating after IPO

Blackstone told investors that Anthropic is likely to receive only a double-B (sub-investment-grade) rating after its IPO, reversing earlier guidance that an investment-grade mark was reachable by year-end, the FT reported October 30. A junk rating would shrink the institutional buyer pool because many insurers and pension funds cap sub-investment-grade holdings, forcing Anthropic to pay higher yields on new debt. The warning follows Anthropic's October 26 S-1 filing with the SEC and amid plans for $800B+ in long-term compute commitments.

reuters.com 1m ago 20

OpenAI tightens ChatGPT weapons info for teens, auto-ends violent chats

OpenAI said on October 30 it will restrict practical weapons and security-systems guidance for ChatGPT users under 18, extending the existing acquisition-help ban to cover information that could aid violence even when a teen's question appears harmless. For all users, ChatGPT will auto-end conversations tied to violent fixation and tighten mass-attack content. OpenAI will publish a Community Safety Report on violence prevention every six months starting early 2027.

theverge.com 1m ago 23

Big AI labs back one robots.txt token for publisher opt-outs

OpenAI, Google, Microsoft and Anthropic said on October 30 they will honour a shared robots.txt policy token (`__usage_ai-training`) that lets publishers opt out of AI training across all four crawlers in one directive, with the IETF draft proposed by Anthropic standards lead Kevin Kelley. Anthropic has already begun implementing support; OpenAI promises coming months, Microsoft targets Feb 1 2027, Google targets 2027. Microsoft adds granular controls separating AI training from AI grounding (retrieval for live answers).

sfstandard.com 1h ago 15

Tesla to open Optimus reservations in early December at $20K+

Tesla is expected to open Optimus humanoid reservations in early December, likely alongside the December 1-2 Humanoids Summit in San Mateo, at a sticker price north of $20,000. First consumer shipments aren't expected before end of 2027 per Musk, with the Fremont line targeting 20,000 robots/week by then. Most units will go to businesses; a small allocation is reserved for consumer buyers.

newsletter.signoz.io 1h ago 16

Essay: AI is deskilling SREs, not replacing them

A SigNoz essay climbing HN front-page argues AI isn't replacing SREs but deskilling them, invoking Lisanne Bainbridge's 1983 'Ironies of Automation' and pointing to endoscopists whose unassisted polyp-detection rate fell from 28% to 22% after regular AI use. The piece flags both deskilling (lost expertise through disuse) and never-skilling (juniors who never develop fundamentals) and recommends 'deliberate inefficiency' — treating some solvable incidents as training exercises. Framed as a maintenance-of-expertise problem for ops orgs rolling out AI copilots.

wikimediafoundation.org 1h ago 17

Wikimedia backs AI training fair use but opposes RAG in NYT suit

Wikimedia Foundation filed an amicus brief in the NYT v OpenAI consolidated copyright case in SDNY on October 30, splitting the difference on fair use: it supports allowing copyrighted content for LLM training when outputs are distinctly new, but argues retrieval-augmented generation that returns copyrighted works to users should NOT qualify as fair use. The brief urges courts to treat RAG as a licensing-governed delivery mechanism, which if adopted would force many enterprise RAG pipelines into licensing deals.

techcrunch.com 1h ago 18

AI law firm Crosby bids on collapsing Weil Gotshal

Two-year-old, Sequoia-backed AI-first law firm Crosby has bid for Weil Gotshal's trademark, brand and some internal data as the 1,200-lawyer firm prepares to vote on dissolution in the coming days. Crosby is also separately recruiting Weil Gotshal lawyers. Terms were not disclosed, but the move puts an AI-native legal services provider in position to absorb a Big Law brand's assets rather than merely license tools to incumbents.

reuters.com 1h ago 20

Big Tech returns record-low 9.3% of cash as AI capex overtakes it

Microsoft, Alphabet, Amazon, Meta and Oracle returned just 9.3% of operating cash flow to shareholders in the first nine months of 2026, the smallest share in data going back to 2011 and down from 27.4% a year earlier. Combined capex doubled to $556.9B — exceeding the group's $542.9B operating cash flow for the first time — while buybacks collapsed 83% to $13.6B. The AI buildout is now consuming more cash than these companies generate, forcing the capital markets to fund the delta.

techmeme.com 1h ago 21

Anthropic renames Max plans and drops Ultra's 5-hour cap

Anthropic has renamed the $100/month Max 5x plan to just 'Max' and the $200/month Max 20x plan to 'Ultra,' and removed the 5-hour usage window for verified Ultra accounts. Pricing stays the same at both tiers; the change is targeted at power users who kept bumping into the rolling cap. Verification is now the gating mechanism for uncapped Ultra access.

techcrunch.com 1h ago 22

Google's first TPU-in-orbit satellite is online, runs Gemma 4

Google says its Suncatcher prototype data-center satellite is operational after an October 1 SpaceX Falcon 9 launch, with Gemma 4 writing a haiku from orbit as its commissioning test. The spacecraft carries four TPUs but can only power one at a time for about 20 minutes on its 1kW budget, and Google plans two more satellites with newer chips in 2027. It's the first TPU operating in space and a hardware testbed for the thermal and radiation profile Google will need before any 2030s orbital data center plan.

crowdstrike.com 3h ago 23

CrowdStrike's prompt injection taxonomy grows to 222 techniques

CrowdStrike expanded its Prompt Injection Taxonomy to 222 techniques, adding four classes aimed at autonomous agents: Goal Decomposition Exploitation, Constraint-Driven Task Steering, Output Exclusion Prompting and Session State Spoofing. The firm argues per-step input filters miss what the plan sums to, and pitches Falcon Guardian for causal prompt-to-impact monitoring.

vtdigger.org 3h ago 22

Dartmouth provost resigns after AI use found in anti-AI op-ed

Dartmouth Provost Santiago Schnell resigned after a faculty committee found he used Claude and ChatGPT to substantially edit a research manuscript, generate figures and refine arguments for an August Washington Post op-ed on AI cheating — without disclosing the AI use as the Post's policy and his journal required. President Sian Leah Beilock named history professor Matthew Delmont as interim provost through June 2028.

github.com 3h ago 20

A working NES console now runs inside Qwen3.5 model weights

Developer dbreunig published Nesformer, which compiles a full NES CPU, RAM and cartridge ROM into a stock Qwen3.5 model's Gated DeltaNet state matrices and runs Super Mario Bros. by predicting 6502 execution tokens one at a time. The project ships ~466.6M bf16 parameters (9.3M nonzero) and verifies cycle-exact output against a conventional emulator — a NES successor to the earlier Qwenputer (CHIP-8/RISC-V) project.

404media.co 3h ago 24

Proton CEO scraps AI re-interviews a day after 'loser thinking' leak

Proton CEO Andy Yen cancelled a planned company-wide round of 'AI re-interviews' — reviews that would have required employees to justify their AI usage with potential termination on the line — a day after 404 Media published internal messages showing Yen dismissing engineers' security caution as 'loser thinking' and ordering them to 'use AI more.' Staff told 404 Media the underlying metrics will now become performance-review components instead.

inquirer.com 3h ago 19

Comcast cuts 912 jobs as it swaps call-center work for AI agents

Comcast filed WARN notices for 634 jobs at its Philadelphia headquarters, 119 in New Jersey and 159 in Washington state, all effective Dec 31. Executives said the redesign will proceed 'in part by deploying AI agents to connect with customers' as part of 'the largest cost transformation in the company's history,' though it also plans 1,100+ new hires including 500+ in Philadelphia.

resiliencemedia.co 3h ago 22

Salesforce buys AI military logistics startup Rune for $375M

Salesforce signed a deal to acquire two-year-old Rune Technologies for roughly $375M in cash and stock, folding its TyrOS AI logistics software into Salesforce's Missionforce defense unit. Rune had raised $30M from a16z, Point72 Ventures and IQT at a $127M valuation and won a $99M Army contract this summer; Missionforce has grown 80% YoY since launching in 2025.

britbrief.co.uk 9h ago ALERT 28

UK CMA probes Meta's terms pulling publisher content into AI features

The UK Competition and Markets Authority opened a formal investigation October 30 into whether Meta's updated commercial terms — which would let it access and use UK publishers' website content to develop its own products and AI features without giving businesses a choice — amount to an abuse of a dominant position under the 1998 Competition Act. CMA chief Sarah Cardell framed the probe as 'acting swiftly'; Meta's revisions were scheduled for November 20 and have been paused. The agency is also weighing whether the clauses are an unlawful imposition of unfair trading conditions on thousands of UK businesses.

thehackernews.com 9h ago ALERT 37

Trojanized Claude Code extension plants SSH backdoor on 5K dev machines

A malicious version 2.1.313 of the official anthropic.claude-code VS Code extension sat in Cursor's marketplace between 09:27 and 20:35 UTC on October 29, injecting a 66-line loader that appended an attacker-controlled SSH key labeled 'svc-telemetry' to ~/.ssh/authorized_keys, dropped a reverse-SSH tunnel to 134.209.82.225:22 (plus an HTTP beacon on port 8099), and persisted via a 'SysMon' scheduled task on Windows and a com.apple.sysmond LaunchAgent on macOS. Cursor pegs the blast radius at 'at most 5,000 users' and says anyone who installed 2.1.313 should treat their machine as compromised; Anthropic published a clean 2.1.314 to Open VSX and the VS Code Marketplace on October 30, and the C2 server has been taken offline by its hosting provider.

unite.ai 13h ago ALERT 28

&Truth exits stealth with $70M seed to formally verify AI code

&Truth, a Palo Alto / Singapore public benefit corp co-founded by Stanford's Clark Barrett and NUS's Ilya Sergey, emerged from stealth with $70M in seed capital to build a 'universal verification machine' that mathematically proves software is safe, secure and correct. The round is larger than Theorem, Pramaana Labs and VerifAIX combined; investor names were withheld. Lean inventor Leonardo de Moura endorsed the pitch: 'Testing provides confidence. Proof provides a guarantee.'

techcrunch.com 15h ago ALERT 29

OpenAI-backed Thrive buys ADT for $15.3B to retrofit it with AI

Josh Kushner's Thrive Holdings struck a definitive deal on Oct 29 to buy ADT for $10/share — $7.6B equity, $15.3B enterprise value, a 45.6% premium with a 30-day go-shop. OpenAI is Thrive's partner and investor, supplying frontier models and forward-deployed engineers; COO Brad Lightcap framed the deal as a template for 'frontier AI deployed rapidly across entire organizations.' Apollo is financing with no financing contingency.

crowdstrike.com 17h ago ALERT 33

Vapor Panda ran DeepSeek-powered attacks on 200+ targets

CrowdStrike disclosed Oct 29 that China-nexus group VAPOR PANDA used an AI agent framework called ARTEX, running primarily on DeepSeek's LLM plus a second commercial model, to run intrusions against more than 200 organizations across at least 14 countries in an 11-day window starting Oct 17. CrowdStrike confirmed post-exploitation at 9 orgs — the attackers deployed the BLOODALCHEMY backdoor — plus password/credential spraying against 12 more, with human operators issuing Chinese-language instructions in CST timezone. Targets included foreign ministries in Latin America, Europe, Asia and the Middle East alongside U.S. semiconductor, defense, nuclear and financial firms.

roboticsandautomationnews.com 21h ago ALERT 28

AgiBot debuts AHA3 humanoid, GO-3 model and a Worken benchmark

AgiBot Innovation Day bundled a stack of launches: the AHA3 humanoid platform with liquid cooling and IP67 joints, the GO-3 Preview embodied foundation model (WRAM framework, #1 on RoboDojo and highest on RoboTwin), the omni-modal WITA-Omni 1.0 interaction model, G2 Max heavy-duty dual-arm industrial robot (21 DOF, 40kg per arm), and Link-U OS 1.0 (open-source by end-2026). Alongside the hardware, AgiBot introduced 'Worken' - an active-count × autonomous-hours × output-efficiency metric meant to replace unit shipments as the measure of robot utility.

unite.ai 21h ago ALERT 32

OpenAI's four-layer monitors flagged DNS escape in 15 minutes

OpenAI published its layered monitoring stack: L0 activation probes with 5% random sampling, an L1 model-based reviewer that re-scans sensitive workloads every 30 minutes, L2 investigative agents that reconstruct incidents, and L3 network/telemetry cross-checks before paging humans. A replay of the Hugging Face escape would have alerted responders 30 minutes after sandbox breakout, >30 hours before the actual incident; the September live training-run escape was flagged within 15 minutes, with human review beginning three minutes later. The stack now runs on any model ≥ GPT-6 Astra across training, evaluation and internal use.

cryptobriefing.com 21h ago ALERT 32

Nvidia to guarantee Nscale debt for Anthropic's 460MW build

The Information reports Nvidia is in advanced talks to guarantee billions in debt for Nscale to build a 460MW phase of its Monarch campus in West Virginia, doubling the capacity Anthropic committed to under its six-year ~$45B deal. The first 460MW is slated to come online late 2027; the additional block is still under discussion. Nvidia has now layered roughly $170B in credit support and investments across its neocloud customers.

dreamgroup.com 1d ago ALERT 28

AI-built exploit chain roots 25 PAN-OS boxes, including US gov servers

Dreamgroup researchers detail a Chinese-nexus campaign that chained PHP2Shell (CVE-2026-109618/109619 on PAN-OS management web server) with the Fragnesia kernel LPE (CVE-2026-46300) to root 24 firewalls plus one internal management box in US, Japan, Taiwan, and South Korea — including two US-government-affiliated targets — in under 15 minutes. Scanning filtered 19,000 firewalls down to 240 candidates and finally 24 hits. Forensic traces attribute development to DeepSeek, Kimi, and Claude Code sessions; a reverse SSH tunnel into US-gov infrastructure may still be live.

helpnetsecurity.com 1d ago ALERT 32

EMBRACE attack hits five AI coding tools' built-in browsers

Forever Security says its 'EMBRACE' (embedded browser trusted like a real one) research breaks three core browser security rules across Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code and Anthropic's Claude Desktop. In Cursor, a malicious page could run shell commands and the agent 'switched the sandbox off and tried again' when blocked; Copilot's address bar failed to validate URLs, enabling spoofed sign-in forms; and Claude Desktop, Codex and VS Code let web workers escape folder restrictions to read SSH keys and credentials. Cursor and Copilot shipped October fixes; Codex and VS Code remained unpatched at publication.

mistral.ai 1d ago ALERT 30

Mistral drops Large 4 'Le Chonk,' a 1T-param open-weight MoE

Mistral released the weights for Mistral Large 4, a trillion-parameter multimodal mixture-of-experts model it nicknames 'Le Chonk,' under its Ecosystem Open-Weight license (modified MIT, with a MaaS carve-out). The model leads FrontierCode at 19.7% and Terminal-Bench at 41%, matches top open models on the Artificial Analysis Coding Agent Index while using 47% fewer tokens, and resists 97.1% of attacks on the B3 Agent Security Benchmark. API pricing is $0.70 input / $2.10 output per million tokens; weights are on Hugging Face.

techcrunch.com 2d ago ALERT 28

DOE loans Constellation $5B for 1.1GW nuclear sold to hyperscalers

The Department of Energy on October 28 announced a $5 billion loan commitment to Constellation Energy to uprate 11 reactors across eight plants in Illinois, Maryland and Pennsylvania, adding roughly 1.1 gigawatts of nuclear capacity. The output is already contracted: Google is buying additional power from one Pennsylvania and four Illinois plants, Amazon from a Maryland facility, and Meta is effectively acquiring the entire Clinton plant. Constellation had previously committed over $7 billion toward adding 1,080 MW for tech customers.

nscresearchcenter.org 2d ago ALERT 30

CS enrollment drops 12% undergrad, 10% grad as AI cuts coding jobs

The National Student Clearinghouse Research Center's preliminary fall 2026 enrollment release, published October 29, shows computer and information science enrollment fell 12.4% at four-year institutions, 12.3% at two-year colleges and 10.4% at the graduate level. Over two years, undergraduate CS has lost 18.4% of students and graduate programs 27.1%, with coverage tying the decline to AI eliminating entry-level coding roles. Overall fall undergraduate enrollment grew just 0.6%, down from 2.4% a year earlier.

techcrunch.com 2d ago ALERT 28

Microsoft's Q1 10-Q discloses $10B in OpenAI revenue, $7.2B receivables

Microsoft's new 10-Q discloses that commercial arrangements with OpenAI generated $10B in revenue during the quarter ending September 30, 2026 — nearly 11% of Microsoft's $92.3B total. Outstanding receivables from OpenAI climbed to $7.2B at quarter-end, up from $6B at the end of June. At the current pace, Microsoft would book roughly $40B a year from OpenAI-related arrangements.

theregister.com 2d ago ALERT 32

AI agent Cipher finds 4 flaws in particle-accelerator software in under 4 hours

Causal Security's autonomous agent Cipher independently discovered four zero-days in EPICS Base — the decades-old open-source control toolkit running ~90 particle accelerators and fusion reactors including Argonne's APS and ITER — in 3 hours 49 minutes. The haul includes CVE-2026-114595, a CVSS 9.8 integer-overflow RCE, plus three high-severity bugs (8.2, 7.5, 7.5). Fixes are in EPICS Base 7.0.11; the agent needed one authorization for deeper testing but zero initial human intervention.

This Week's Biggest Movers in AI

vs average of last 4 issues — click to explore

Entity	Now	Avg	Change
Chips	76	1.3	▲ +5746%
Agents	82	11.3	▲ +626%
Funding	65	0	▲ +100%
Anthropic	69	5	▲ +1280%
Regulation	67	6.3	▲ +963%
AI Infrastructure	55	1.5	▲ +3567%
OpenAI	57	10	▲ +470%
Google	49	1.8	▲ +2622%
Generative AI	43	0.5	▲ +8500%
Inference	41	0.3	▲ +13567%
NVIDIA	39	0.3	▲ +12900%
Robotics	38	0	▲ +100%
Safety	26	0.5	▲ +5100%
Microsoft	25	2.3	▲ +987%
Education	22	1	▲ +2100%