Today in AI — November 3, 2026
Apple shipped a Siri update and a new risk factor on the same day, OpenAI spent the day hiring people instead of GPUs, and two memory supply chains bent around export controls. The day's quieter theme: the model is no longer the product — the licence, the proof and the permission are.
Models & Research
OpenAI says GPT-6 Astra designed a hardware device, end to end, in under a week. According to OpenAI, its hardware team used the model to design a device for talking to the Codex coding agent — enclosure, circuit board layout with component selection and trace routing, firmware and interface — and a contract manufacturer produced a small batch of prototypes in under a month, after several revision rounds when the model failed to account for manufacturing tolerances and assembly details.
The number that undercuts the pitch is in the same post: on BenchCAD 2.0, which tests whether a model can reconstruct 3D parts and trace connections on a board, Astra at its highest reasoning setting scores 0.594, while an unnamed "internal model" scores 0.875. The design capability being demonstrated is not quite the capability being sold — and with token use on OpenAI's hardware teams up 23 times since February, the next version of this claim will be a cost argument, not a capability one.
Anthropic open-sourced a way to prove which model answered you. FlashProve, from Anthropic's Frontier Red Team, produces a zero-knowledge proof that the tokens a user received came from the exact weights the provider committed to, without revealing those weights: about 110 tokens per second on a single H100 for a 400-billion-parameter mixture-of-experts model, with 19 MB proofs that a CPU verifies in seconds. That is still 70 to 95 times the GPU time of ordinary inference, and the proofs cover greedy decoding over 1,024-token windows, so nobody is serving production traffic with it.
The reason to care is regulatory rather than commercial. Proof-of-inference is the missing piece in any regime that wants to verify a lab is really serving the model it claims — or that it is honouring a compute commitment — and Anthropic's own frontier safety roadmap set a prototype deadline of September 30. Formal verification in Lean is the other half of the release: the security claims are machine-checked, which matters more for cryptography than for model behaviour.
Industry
OpenAI's capital is now hiring people, not just compute. OpenAI has hired Jill Hazelbaker, Uber's president and chief corporate affairs officer, to run corporate affairs and report to Greg Brockman, as Axios first reported and Bloomberg confirmed; Bloomberg notes she is the latest senior Uber executive to leave this year. The same day, the Financial Times reported OpenAI hired former Army Secretary Dan Driscoll as senior counsellor for national security modernisation.
Two hires, one direction. Communications and defense are the functions that decide whether a lab gets treated as infrastructure or as a defendant, and OpenAI is staffing both at the level of people who have run large institutions — not the policy-shop hires of two years ago.
SK Hynix's New York fab plan is stuck behind a licence for its Chinese one. Reuters reported that US officials have held up approval for American chipmaking equipment SK Hynix needs to expand at Dalian, which the company argues would generate the revenue to fund its US buildout. Its Solidigm unit plans to invest about $45 billion in two NAND plants in Marcy, New York, potentially $100 billion over time; executives have met Commerce officials nearly ten times since late August, and Under Secretary Jeffrey Kessler has held to the position that no new tools go to upgrade or expand Chinese capacity.
The trap is structural. Controls written to keep equipment out of China are now the gating item for bringing memory production onshore, and the county executive for the site says a federal investment tax credit expires at the end of the year — turning a national-security review into a race against a tax deadline.
CrowdStrike is buying the agent-security layer instead of building it. TechCrunch reported CrowdStrike will pay roughly $495 million for Fiddler AI, which monitors AI agents inside enterprises, and Dreadnode, which stress-tests models and agents before deployment — about $400 million and $95 million, mostly cash — with both expected to land in the Falcon platform early next year. The same day it launched an AI kill switch in Falcon Guardian to stop internal agents that go rogue, and announced a Falcon Guardian integration with Anthropic's Claude Cowork.
Security vendors spent 2026 shipping agent features. This is the first time one has paid half a billion dollars for the observability and red-teaming to back them up, which tells you where the buyers' anxiety actually is: not whether agents work, but whether anyone can prove what they did.
Russia is moving its data centers underground after Ukrainian drone strikes. Rostelecom has discussed putting data centers in Norilsk's mine workings, and a regional governor has proposed disused open-pit coal mines with public parks planted on top as camouflage, Ars Technica reported. On November 2, Vladimir Putin signed a decree reserving 10 percent of commercial data center capacity for state needs and allowing a security commission to authorise foreign data centers — a reversal for a country that built a "sovereign internet" to keep compute inside its borders. Yandex has already relocated all three of its main supercomputers.
The comparison is uncomfortable for everyone else. Russia's hyper-centralised cloud was cheap to build and easy to map; Ukraine's dispersed, partly cloud-migrated services have been the more resilient system. Concentration is a single point of failure the moment the threat is physical, and that argument now applies to any country that has bet its AI strategy on a handful of campuses.
Policy
Britain's cyber agency put a date on AI-powered ransomware. The NCSC's annual review says AI-powered malware is "highly likely" to be used in a ransomware attack on UK critical national infrastructure by 2028, and its chief executive, Richard Horne, puts the cause in underinvestment rather than novel technique: vulnerabilities "tolerated for far too long" can now be found and exploited at a scale that was once unimaginable.
The report lands alongside money announced last week — £60 million for a first phase of a national Cyber Shield, £70 million for a vulnerability centre, and an AI security centre of excellence in Manchester — but Horne's useful sentence is the unfunded one: Britain has "got away with doing tech on the cheap". That is an argument about hygiene, not model capability, and it is the same failure mode we covered when Anthropic said lone hackers now run state-level cyber campaigns.
Apple added AI agents to its annual risk factors. In the 10-K it filed Tuesday, Apple warned that agents "may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized," and that its "ability to oversee and control third-party artificial intelligence agents and other technologies may be limited," opening the company to lawsuits and reputational damage.
Apple does not ship an autonomous agent, which is the point: its exposure is the agents users install on Macs, the same surface it moved to restrict when it tightened full disk access last month. From the company that has sold privacy as its differentiator, the filing is an admission that the operating system is now the enforcement point for someone else's autonomy — and that Apple knows it cannot fully be one.
Tools
Siri AI now takes actions inside any app that adopts App Intents. With the iOS, iPadOS, macOS and visionOS 27.3 developer previews, Apple opened Siri AI to all developers: apps can hand the assistant their content and actions, users can ask for things without naming the app while it is on screen or in a Live Activity, and a new "app skills" feature chains several tasks into a single request.
Apple's own apps get the same treatment across Mail, Photos, Notes, Safari, Files, Maps and others — find the coffee shop you visited last week, archive files untouched for a year. This is the part of the Siri rebuild that decides whether it is a feature or a platform: an assistant is only as capable as the number of apps willing to grant it verbs, and for the first time third-party developers have a reason to.
The thread through today is that permission is becoming the scarce resource — a licence for a fab, a proof for an inference, a kill switch for an agent, an API for a phone. Where do you think the real bottleneck sits? Write to us via our contact page.
