Afternoon Brief · October 31st 2026

Afternoon Brief, October 31st 2026

Apple Intelligence is coming to China, and Alibaba's Qwen will answer part of Siri there. Anthropic is now expected to get a junk rating after it lists. A new asks Claude what each victim is worth. And Meta has quietly built a with a Chinese robot maker.

13 min read·Editorial by Nour Haddad

NHNour Haddad
Deployment & the Floor

Apple Intelligence reaches China in December, with Alibaba's Qwen answering part of Siri

Apple's China newsroom confirmed Friday that Apple Intelligence testing in mainland China starts November 4th in the next iOS developer beta, with a public launch in December alongside iOS 27.3. Apple says it has partnered with Alibaba to add Qwen to parts of the system: some requests in the improved Siri, Writing Tools and Visual Intelligence. The first wave includes Live Translation, Genmoji, Clean Up in Photos, natural-language photo search, notification summaries and Smart Reply, on the iPhone 15 Pro and newer, with Siri set to English or Simplified Chinese.

What is missing is most of what Apple shipped this year. The new Siri AI is excluded, and Macs and iPads in China wait for further regulatory approval. Apple Intelligence launched in the US two years ago; in its third-largest market it arrives with a Chinese model doing part of the reasoning, and every future model update there needs its own sign-off from Beijing. SourcesB B C

Blackstone now expects Anthropic's credit to be rated junk after its IPO

Blackstone has told investors in the $63 billion chip-financing deal it anchored for Anthropic that it now expects a double-B rating once Anthropic lists, the top tier of speculative grade, the Financial Times reported Friday. Blackstone had previously expected Anthropic to reach before 2027. Anthropic and Blackstone declined to comment.

A double-B rating does not stop Anthropic borrowing, but it raises the price of every dollar it borrows, for a company that has signed some of the largest compute contracts in the industry this year. It arrives in the same week S&P put Meta, an AA- credit, on for its spending. SourcesB

A new infostealer sends each victim's data to Claude and asks what they are worth

Cisco Talos on Saturday described VOLTA, an infostealer that hands its haul to a language model for appraisal. After stealing , browser sessions, cards and crypto wallets, it builds a profile of the victim and sends it through to Claude Opus 4.6 at temperature 0.2, with a system prompt casting the model as "a forensic intelligence analyst drafting a target assessment." If the reply contains a refusal phrase such as "I can't," VOLTA retries on Moonshot's Kimi K2, then Google's Gemini 2.5 Flash.

The required output includes a section headed "WHAT THEY'RE WORTH," ranking the victim's assets into low, medium and high tiers by how easily they can be cashed out, and another assessing the victim's password habits. VOLTA also uses the stolen sessions to fill gaps, reading Gmail receipts to confirm a home address and pricing Roblox inventories. Talos first saw a sample on October 11th. The fallback chain is the detail defenders should take from it: a refusal from one provider costs the attacker a single retry on another. SourcesA

OpenAI opens GPT-Rosalind-6 to vetted labs, with fewer refusals on dual-use biology

OpenAI published a safety addendum Friday for GPT-Rosalind-6, a biology and drug-discovery model offered only through its trusted-access program to vetted scientists, institutes, companies and government biodefense teams. It uses the same as GPT-6 Astra with no further training, configured to refuse fewer biology questions. OpenAI rates it High in biological and chemical capability, below Critical, and Critical for cybersecurity.

Against GPT-Rosalind-5.5 it scores 49.6% against 28.9% on MedChemBench, 37.7% against 16.7% on GeneBench Pro and 61.3% against 44.0% on LifeSciBench, all OpenAI's figures. None of its four Critical-level biology evaluations crossed its threshold. Approved customers must restrict access to named users and have qualified people review outputs before they shape experiments or publications. SourcesA

Meta built a humanoid robot with China's Fourier in about three months

Meta has built a prototype humanoid called Max with the Shanghai robot maker Fourier and is working on a second version, The Information reported Saturday. The robot weighs about 106 pounds and was shown to some Meta employees at an internal robotics summit in late August. Meta's robotics team worked at Fourier's Shanghai factory, with the company's vice president of product for robotics, Li-Chen Miller, spending almost two months there, according to summaries of the report.

Meta reportedly sees the humanoid as a body for its Muse and wants it doing household tasks, with no plan to sell the units. A US company building its robot body in Shanghai lands three days after the White House declared robot imports a national security threat. Reuters could not immediately verify the report. SourcesB B C

Baseten seeks about $2 billion at up to $30 billion, more than double its June price

The company Baseten is raising about $2 billion in two , the first at a $26 billion and the second at $30 billion, with Sands Capital expected to lead, Bloomberg reported Friday. Terms are not final and Baseten has not announced the round. The top price is 2.3 times the $13 billion Baseten was valued at in June, when it raised $1.5 billion.

Inference is the part of the AI bill that grows with use, and investors are paying up for it in the same week OpenAI's round shrank by a third. SourcesC C

Built-in browsers in five AI coding tools could reach the files on a developer's machine

Researchers at Forever Security reported that the browsers embedded in Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code and Anthropic's Claude Desktop were trusted like the computer behind them. In Cursor a web page got the agent to run a hidden command. In Copilot a page could show a genuine github.com/login address over a fake sign-in form. In Codex, VS Code and Claude Desktop a single local file, opened with one click on a link the assistant wrote, could read keys, .env files and off the disk. Forever calls the pattern EMBRACE.

The vendor record is uneven. Cursor patched on October 1st, three weeks after disclosure, without replying. GitHub fixed it in Copilot 1.1.29 before a report was filed. Anthropic declined Forever's report but accepted a near-identical one from Noma Security a day later. Microsoft declined, citing Workspace Trust, and OpenAI had not replied since July 27th; neither had a fix as of Thursday. SourcesB

Sebastian Thrun's Dulo raises a $130 million seed to build foundation models for hardware design

Dulo, the San Francisco startup founded by Sebastian Thrun and Adrian Aoun, has raised $130 million in seed funding from investors including Eric Schmidt, John Doerr and DST Global, The Information reported Friday. Aoun describes the goal as "building for hardware design": automating engineering design and validation, and leaving manufacturing to others. Thrun, who led Stanford's AI lab and Google's self-driving project, first showed Dulo at the Actuate conference in August.

No valuation, lead investor, customers or product specifications have been disclosed. It joins &Truth's $70 million seed this week in a run of nine-figure first rounds for teams applying models to engineering work. SourcesB C

All 79 of Victoria's councils call for a halt to new data centers

The Municipal Association of Victoria, which represents every council in the Australian state, passed a resolution calling for a moratorium on approving new or expanded data centers until a full planning and environmental framework exists, ABC News reported Friday. Eight councils brought motions, including Melbourne and Geelong. The association objected to a state program that lets data-center companies apply directly to the planning minister, bypassing councils, and pointed to NextDC's purchase of 169 hectares at Lovely Banks in Geelong, land earmarked for more than 2,000 homes.

The timing limits what it can change. Victoria's government enters caretaker mode in early November ahead of a state election, and its September rules already ban data centers near homes and schools and require them to bring their own renewable power. SourcesB

South Australians rally against a proposed 1,200-megawatt data center in the Adelaide Hills

Residents rallied in Mannum on Saturday against Quantum DC's proposed 1,200-megawatt data center at Tungkillo, about 41 kilometers east of Adelaide, on a 120-hectare site that would include a gas-fired power station. No development application has been lodged, and Quantum DC has pushed its community consultations to November 20th and 21st, citing safety concerns. "We will fight it using every legal means possible," organizer Magdalene Lippett said.

South Australia's planning minister said any application would have to meet power and water requirements. The Greens want a moratorium; the Liberal upper-house leader said her party is "not against data centres." Two Australian states produced organized opposition to data centers on consecutive days. SourcesB

Apple acquihires Ascend.io, an agentic data-engineering startup, through a filing it had to make in Europe

Apple has struck a deal for Ascension Labs, better known as Ascend.io, which sells AI agents that build and maintain data pipelines, according to the EU's public list of acquisitions under the . The filing gives Apple the right to hire certain Ascend employees, a non-exclusive license to its intellectual property and possible assignment of some patents. No price was disclosed.

Apple seldom announces deals this size; it surfaced because the DMA requires gatekeepers to report acquisitions. Ascend was founded in 2015. SourcesB

Moonshot closes its Kimi K2 models to new users and will retire them by the end of December

From today, Moonshot no longer offers its Kimi K2 series to newly registered accounts, and it plans to retire all K2 models by the end of December, its developer documentation says. Existing users are unaffected for now and are pointed to Kimi K3.1 Pro. K2 was the model that put Moonshot on the map in July 2025; the company is retiring it from its API while its successor's weights are still promised by November 4th. SourcesA C

GPT-5.6 Sol launched a rocket in Factorio, playing under a human's limits

A developer, Ryan Madden, reported Saturday that GPT-5.6 Sol, running in Codex at a low thinking setting, finished a peaceful game of Factorio (no enemies) and launched a rocket after 13.5 hours of game time. A held the model to the limits a human player faces through the interface, including reach and fog of war. The model never connected some late-game production lines and carried supplies between factory areas by hand instead. GPT-5.5 never used conveyor belts in the same setup. One run, one person's harness, and no enemies. SourcesA

Sources

40 citations · 8 primary · 20 · 12 weaker

Morning Brief, October 31st 2026