AI today, 7 November 2026: DeepMind weakens old RSA, METR flags Anthropic's speed-up, data centres hit the ballot
12 things that happened in AI, each in under a minute. Today's news is about limits and speed. DeepMind's agents chipped at an old cryptographic standard, and METR says AI is now doing much of the work of speeding up AI itself. Meanwhile voters, investors and prosecutors pushed back on the physical and legal footing of the build-out, and Google and OpenAI responded with power and community pledges.
-
Research · Big
DeepMind says AI agents found two weaknesses in legacy 1024-bit RSA within weeks
Google DeepMind's Pushmeet Kohli announced on 6 November that AI agents, under human supervision, found two advances against legacy 1024-bit RSA, Interesting Engineering reports. One is a signature-forgery technique said to be up to 1,600x faster than previous methods; the other reportedly cuts the average cost of factoring each additional key in a batch by about 30x. The article says modern encryption in use today is unaffected, since 1024-bit keys have largely been replaced. The figures come from DeepMind's own announcement, not independent checks.Why it matters: It is a concrete case of AI helping to weaken a long-trusted cryptographic scheme; anyone still running 1024-bit RSA keys should plan to retire them.
-
Safety · Big
METR: Anthropic's AI capability progress is about twice its earlier pace, mostly AI-driven
In a report published 5 November, the independent evaluator METR estimated that Anthropic's benchmark scores have grown about 1.9x faster since October 2025 (95% interval 1.5–2.8x). It judges, subjectively and with high uncertainty, that about 80% of the October 2026 acceleration is due to AI assistance, and that faster training compute is not the cause. Anthropic redacted a majority of the evidence METR found most informative, so readers must lean on METR's judgment more than is ideal.Why it matters: It is the clearest outside estimate yet that AI is speeding up its own development, and the heavy redactions limit how far anyone can check it.
-
Policy · Big
At least 40 local data-centre curbs pass in US midterms; OpenAI answers with a community compact
At least 40 local laws to curb and scrutinise data-centre construction passed across the US during the midterms, The National reports. On Thursday OpenAI announced a National Community Compact that promises to cover the energy costs of the infrastructure it builds, protect local resources, publish local jobs reports and stop using non-disclosure agreements with public officials, with narrow exceptions. The compact calls data centres "compute campuses".Why it matters: Local opposition is becoming a real limit on AI build-out, and the biggest labs are now changing their pitch to communities.
-
Business
Google commits $17B to Wyoming data centre powered by 702 MW of fuel cells
On 5 November Google announced a $17 billion investment in its first Wyoming data centre, near Cheyenne. Power comes partly from a 702-megawatt natural-gas fuel-cell development (Bloom Energy fuel cells, with carbon capture planned) under a 20-year take-or-pay power deal with AEP, and Google says it will pay for 100% of the power used at the site. Google also met residents who voiced unease about data-centre growth and is forming a community council, per Yahoo Finance.Why it matters: Hyperscalers are building their own power and promising to carry the costs, a direct response to the local backlash.
-
Business
AI stocks fall as investors question whether the infrastructure boom can hold its pace
The Neuron's roundup reports AI stocks fell as investors asked whether the infrastructure build-out would keep its expected pace, pointing to governments in New York and Amsterdam tightening data-centre restrictions. The National also links Thursday's AI and tech stock drop to investor reaction to Democratic gains in the midterms. Exact index moves were not verified.Why it matters: A slower build-out would hit chip, power and cloud suppliers and the labs that have committed to large compute bills.
-
Policy
Justice Department launches a Super Intelligence Task Force to police AI misuse
The DOJ announced the task force on 5 November, set up by an Attorney General memo dated 13 October. It targets foreign governments, criminal groups and actors committing fraud or other unlawful conduct with AI, and will coordinate criminal and civil enforcement and issue prosecution guidance, per Nextgov. It absorbs the earlier AI Litigation Task Force, and Tech Times reports it creates no new crimes.Why it matters: It signals federal prosecutors preparing AI-specific cases, though no priorities have been named. Background: /explained/trump-super-intelligence-force-explained/.
-
Models
Google makes Gemini Omni 1.2 Flash generally available: 15-second video, 30% cheaper
On 5 November Google made its video generation and editing model generally available in the Gemini API and Google Cloud. Clips now run up to 15 seconds (up from 10), and 720p costs $0.07 per second, 30% less than Omni 1.1. It reaches Flow, Vids, Google Ads and the Gemini app (Ultra subscribers first), and outputs carry a SynthID watermark.Why it matters: Cheaper, longer AI video lowers the cost of ad and short-form production for businesses.
-
Open source
Thinking Machines releases Inkling 2 Small via API, promises Apache 2.0 weights next week
The 594B-parameter mixture-of-experts model (17B active per token) takes image input and is aimed at coding and visual work. It is available through Tinker with closed weights while safety testing finishes; the company says full Apache 2.0 weights follow "next week". The company reports 86.3% on SWEBench Pro v2 (Hard), second to Kimi K3's 88.2% in its own table (vendor-reported).Why it matters: If the weights ship as promised, it is a large open model strong at coding that teams can host themselves.
-
Products
Anthropic adds read-only Personal Finance to Claude for US paid users
The beta, announced 5 November, lets US Pro, Max and Ultra subscribers connect bank and credit-card accounts through Plaid and ask about spending, duplicate charges and rising bills. Claude sees up to two years of transactions but cannot move money, per TechCrunch. Anthropic says Personal Finance conversations are not used for training and that there are no ads or sponsored recommendations.Why it matters: Assistants are moving into sensitive personal data; ChatGPT launched similar tools in May.
-
Business
Meta acquires Conductor, a tool that runs several AI coding agents in parallel
Meta confirmed the acquisition to TechCrunch on 4 November; terms were not disclosed. Conductor works with agents from Anthropic, OpenAI and Cursor. Its cloud service shuts down on 20 November, after which user cloud data will be deleted; core features keep working as now until early next year.Why it matters: Big labs are buying the tooling around coding agents, and users of Conductor's cloud service have about two weeks to move.
-
Open source
Baidu releases ERNIE 5.5 Flash Base, an open 280B-parameter foundation model
The Decoder reports the model has 280 billion parameters, 14 billion active per input. Released on 5 November without any post-training so developers can adapt it themselves; weights are on Hugging Face and a post-trained version is expected later.Why it matters: Another large open base model from a Chinese lab for teams that want to fine-tune their own assistants.
-
Hardware
Agility Robotics lists on Nasdaq via SPAC at about $2.5B valuation
Trading began 5 November under AGLT after a SPAC merger with Churchill Capital Corp XI, with about $465M in gross proceeds (including about $164M from the trust and a $201M Foxconn-led PIPE), per The Robot Report and company filings. The $2.5B figure is the pre-money valuation; Agility reported only $1.78M in 2025 net sales and a $140.2M operating loss.Why it matters: It is a test of public-market appetite for humanoid robots with real commercial deployments.
How this was made: compiled by an AI model (Claude) from the linked sources and checked item by item in a separate AI fact-check pass. Corrections: [email protected].