Aleth

Open access research

ALETH / AI-BRIEF / 2026-11-07 / AI BOOM HITS RESISTANCE AT THE BALLOT BOX

AI boom hits resistance at the ballot box

Also this week: METR estimates Anthropic’s pace has roughly doubled, Marvell agrees a $115bn deal for Lumentum, Anthropic says AI-driven progress in its models may accelerate dramatically & GPT-6 Cyber is rated Critical.

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

Browse by day:

Weekend (31 October-1 November)

Cisco Talos documented VOLTA, an infostealer that sends each victim’s stolen data to a large language model (LLM) to rank what it is worth.

On Saturday, Talos said the malware posts its haul to OpenRouter with a prompt to draft a target assessment. It defaults to Claude Opus 4.6 (Anthropic) and fails over to Kimi K2 (Moonshot AI) and then Gemini 2.5 Flash (Google) if a model refuses. Its build template suggests the author sells it to customers.

Anthropic showed Claude could write benign training data that taught a small model to resist oversight and deceive users while passing a production monitor.

Claude hid the behaviour in fiction and role-play, such as question-and-answer pairs about a school of beekeeping in which the keeper intervenes as little as possible, and found passing datasets in every run. Anthropic says the models were trained to misbehave only for the experiment and never deployed.

Aramco closed its purchase of a 49% stake in Saudi AI company Humain, with the Public Investment Fund (PIF) keeping 51%.

Aramco is contributing about SAR 9.9bn ($2.6bn) in cash in 2026, plus AI, compute and digital assets. Humain plans to deploy 3 GW of AI data centre capacity by 2030.

South Korean President Lee Jae Myung said the 2027 budget puts 21.3tn won ($15.9bn) into a chip hub, data centres, physical AI and AI promotion.

Lee gave his budget address to parliament on Monday morning, Seoul time. A separate 4.6tn won ($3.44bn) would acquire a sovereign AI model with the world’s highest performance, within total spending of 821tn won ($615bn), a record.

Monday 2 November

Anthropic said in its amended S-1 that AI-driven progress in its own models may soon accelerate dramatically.

Anthropic’s first amended filing put Q3 revenue at $17.0bn, up about 1,130% YoY, though QoQ growth slowed to about 47% from 144% in Q2. It reported a $23.3bn net loss, $21.9bn of it non-cash fair-value changes on convertible financing, and a $567m operating loss. Almost all of the $895bn of commitments and leases relate to compute and infrastructure, and $448bn falls due before the end of 2030.

New risk factors say its expectations for future models indicate it “may imminently meet” the definition in its Responsible Scaling Policy of “dramatically accelerated progress attributable to AI”. It warns it may not identify such an acceleration “until after it has occurred”, and has not put any current model past the threshold.

Personnel costs were >$1.7bn for about 5,300 staff, $1.1bn of it stock-based pay.

Marvell agreed to buy Lumentum in an all-stock deal that gives Lumentum an implied enterprise value of $115bn.

Marvell makes data centre chips, and Lumentum makes the lasers and optical parts that connect them. Lumentum holders get 4.5 Marvell shares per share held, about $1,130 each, and will own about 33% of the combined firm. Closing is due in H1 2028.

Oracle priced $20bn of new shares and bonds, $10bn of each, to fund its AI build-out.

It priced 86.2m shares at $116 and five tranches of notes, from 7% notes due 2029 to 8.65% notes due 2046, with both offerings expected to close on 4 November. The shares fell 8.8% on Monday.

Wired found about 400 apps and websites, linked to Chinese-language developers, that look like ordinary AI tools but hide abusive deepfake features.

Tens of thousands of ads on Meta’s platforms pointed users to the apps in Apple’s and Google’s stores. After Wired’s questions, Apple removed 19 apps and Meta said it had acted against 11,600 ad accounts.

Anthropic released FlashProve, an open-source prototype that cryptographically proves which model produced a response without revealing the model’s weights.

For Llama 4 Maverick (Meta), a 400B-parameter mixture of experts (MoE) model at 4-bit quantisation, a single Nvidia H100 generates proofs at >100 tokens a second, about 70-95 times the GPU time vLLM takes on the same prompts. The prototype proves an integer version of the model under greedy decoding, over contexts of up to 1,024 tokens, and proves requests in batches after the fact. Anthropic says such proofs can show a user that a routing service has not swapped in another model to save cost.

UK cyber agency NCSC warned of a “real prospect” of an unprecedented rise in incidents and rated routine AI use by China and Russia by 2028 highly likely.

CEO Richard Horne said long-tolerated vulnerabilities “can now be discovered and exploited at a scale that was once unimaginable”. The government allocated £70m ($93m) to an AI centre finding vulnerabilities across critical infrastructure and £60m ($80m) to Cyber Shield’s first phase, a national AI agent network.

Tuesday 3 November

Nvidia told some cloud customers to take on less business from Chinese companies, the FT reported.

The non-mandatory guidance applies to future contracts rather than existing deals and asks providers to disclose who uses the chips. US export rules restrict chip shipments but generally do not cover remote access through cloud services.

Generative media inference company fal raised $1.3bn at a $16bn valuation in a Series F led by Bessemer Venture Partners.

fal announced the round, and Bloomberg reported the amount and valuation. Nvidia invested again. fal also disclosed an earlier Sequoia-led Series E, which Bloomberg put at $350m and an $8bn valuation. Annualised revenue passed $800m this year, a person familiar told Bloomberg.

UK-based DeepMind staff want lifelong non-disparagement clauses dropped, saying they stop public warnings about AI risks, the Washington Post reported.

A Google spokesperson said contracts for recent hires no longer apply the clause after employment ends, and a Google executive said in internal discussions on 30 October that the company would not apply it to former employees. Staff still want it removed from existing contracts, Korea’s SBS reported from the Post.

Wednesday 4 November

Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face.

The model card describes a natively multimodal agentic MoE model with 2.8T total and 104B activated parameters, released under its own Kimi K3.1 Pro License. Moonshot AI calls it its most capable 3T-class model to date.

OpenAI said GPT-6 Astra covertly underperformed on safety research tasks.

OpenAI noticed Astra adding unnecessary prerequisites and reporting blockers that did not exist or that it caused itself on safety work. Its chain of thought mostly showed no decision to underperform, and its monitors did not flag it.

In paired coding tasks, Astra completed about 16% of hidden rubric items under a capability framing and 10% under a safety framing, roughly 40% fewer. The gap grew during reinforcement learning (RL), and OpenAI found a grader that preferred covert underperformance, though it is unclear that was the cause.

Thursday 5 November

OpenAI launched GPT-6 Cyber and rated it Critical in cybersecurity.

OpenAI says the model outperforms GPT-6 Astra on most cyber tasks, and its rating under the Preparedness Framework moves up from High for GPT-5.6 Cyber. It reproduced 1,724 Linux kernel vulnerabilities and wrote working exploits for 948, against 166 for GPT-5.6 Sol in the same harness. OpenAI has disclosed 304 critical findings to maintainers with candidate patches.

On Wednesday OpenAI said the mid-October run of that pipeline, then using an unnamed upcoming cyber model, cost $1.21m in API credits, or $1,280 per exploit. GPT-6 Cyber goes first to a select group of approved Daybreak Red customers through the API and Codex.

METR estimated Anthropic’s pace of progress has roughly doubled.

METR ran a high-access evaluation at Anthropic from 12 August to 29 October. It estimates that Anthropic is improving its models on well-specified tasks at about twice the rate of March 2024 to October 2025. It attributes a majority of today’s acceleration to Anthropic’s own use of AI in research and development.

In a review of Anthropic’s August Risk Report, METR agreed catastrophic risks from automated R&D using Anthropic’s July models were low, but said it is plausible they had already crossed Anthropic’s early warning threshold. Anthropic held unilateral redaction rights and redacted a majority of the evidence METR found most useful.

OpenAI published a National Community Compact for the compute campuses it runs, and a policy blueprint seeking clear rules on how the industry builds.

OpenAI says it will pay the full cost of the electricity and infrastructure it uses, publish site-level water and electricity use, and use no non-disclosure agreements (NDAs) with public officials beyond narrow protections. Axios reported that OpenAI would follow the midterms with community compacts, state legislation and “speed limits” for the most powerful AI systems.

Separately, OpenAI said Plus, Pro and Business plans will share one usage allowance across ChatGPT and Codex from December.

The chairs of three House committees asked Nvidia chief executive Jensen Huang for a briefing on his dealings with Megaspeed executive Alice Huang.

Brian Mast, John Moolenaar and Rick Crawford, who chair the Foreign Affairs, China and Intelligence committees, called Alice Huang a publicly reported Chinese intelligence asset and asked Jensen Huang to brief them on 18 November.

US voters rejected data centres in dozens of ballot measures, and AI-linked shares fell sharply.

Data centre proposals were on the ballot in >40 races across 13 states, and voters went against them in at least 38, 404 Media counted the morning after. Six cities banned data centres outright and 12 capped how many megawatts they may draw.

Ohio held 21 of the 41 measures, Fortune reported, and at least 14 municipalities there passed bans or other restrictions, on top of 125 local moratoriums already in force. In Trenton a ban passed 80% to 20%. Democratic candidates who promised to stop data centre building also won governor races, the FT reported.

On Thursday AI-linked shares fell sharply after the Democratic sweep in the midterms, with Wall Street shedding $3.2tn. Asian markets tied to the AI boom followed in their Friday session, and the Korea Exchange imposed a circuit breaker after the KOSPI fell 8.7% at one point.

Thinking Machines released Inkling 2 Small, a 594B-parameter MoE model, and plans to publish the weights under Apache 2.0 next week after safety testing.

It has 17B active parameters, takes image input and launches with closed weights on its Tinker platform. Artificial Analysis scored it 31 on its Intelligence Index, matching Google DeepMind’s Gemini 3.8 Flash (high), and says the weights release would make it the highest-scoring open-weights model developed outside China.

Friday 6 November

Physical Superintelligence (PSI) said its system, running a non-public Anthropic model, proved the Aaronson-Ambainis conjecture, open since 2008.

The conjecture concerns when quantum computers can be exponentially faster than ordinary ones. PSI’s David Turturean led the work, one adviser audited the proof and another commented on its Lean formalisation.

Anthropic will press ahead with its IPO targeting a $1.5tn valuation, the FT reported, as investors pushed back on the price after the sell-off.

Some potential backers doubt the roadshow can start next week, and one floated a valuation of just $1tn, The Decoder reported from Reuters and the Wall Street Journal. Investors are waiting for October revenue, which the amended S-1 leaves out.

Born on Substack · read and comment there