The AI Daily Five · No. 21 · Friday 6 November
Who Really Holds Power Over AI?
From Silicon Valley’s midterm losses and Claude’s access to bank data to METR’s assessment of Anthropic, smartwatch screening and tokenised bonds, today’s edition asks who holds power over technology.

1
Did Silicon Valley’s Election Gamble Backfire?
Technology billionaires spent more than $600 million supporting Republicans in the US midterms. Democratic victories and local opposition to data centres have exposed the limits of the industry’s political influence.
Elon Musk spent nearly half a billion dollars supporting Republican candidates in the US midterm elections on 3 November. The results offered little return on that investment: according to Reuters, just three of the 37 Republican House candidates he supported had won by Thursday afternoon, while 33 had lost and one race remained undecided.
Musk was not alone. The Financial Times reported that Silicon Valley billionaires and technology interests spent more than $600 million backing Republicans, including contributions from venture capitalists, cryptocurrency entrepreneurs and other influential figures in the technology industry. Despite that spending, Democrats secured control of both chambers of Congress.
The scale of Musk’s involvement was particularly striking. He directed substantial funding towards Texas, where his companies Tesla and SpaceX have major operations. Reuters reported that he spent almost $230 million on the state’s Senate contest alone, while also supporting candidates in its gubernatorial and lieutenant-governor races.
Several prominent technology-backed candidates lost, including Texas Republican Senate candidate Ken Paxton. The results have renewed attention on the relationship between wealthy technology executives and the political decisions affecting their businesses.
For AI companies, congressional control matters. Federal lawmakers can investigate corporate conduct, scrutinise public contracts and examine questions ranging from AI safety to competition and government procurement. However, winning congressional majorities does not automatically translate into new legislation, particularly where the executive branch and Congress disagree.
Opposition to AI infrastructure also emerged at the local level. According to Ballotpedia’s election review, voters favoured restrictions on data-centre development in 40 of 41 local ballot measures held on 3 November.
Some measures imposed bans or temporary moratoriums, while others strengthened zoning requirements or expressed opposition to proposed developments. Their consequences vary by jurisdiction, but the overall pattern is significant.
Data centres are increasingly central to the expansion plans of AI companies. They also require substantial electricity, water, land and infrastructure, making their development a concern for local communities.
The election results therefore reveal two dimensions of the industry’s political position. Technology companies have considerable financial resources, but voters and local institutions can still challenge their priorities.
The results cannot be attributed exclusively to attitudes towards technology. Economic conditions and wider political issues also influenced the elections. Nevertheless, the industry’s investments failed to secure many of the outcomes its donors supported.
2
Claude Can Read Your Bank Transactions. What Next?
Anthropic’s new personal finance service allows users to connect bank accounts and analyse up to two years of transactions. The feature offers convenience while bringing sensitive financial information into another technology platform.
Claude can now examine users’ bank balances, credit card information and transaction histories through a new personal finance feature introduced by Anthropic on 5 November.
The service, Personal Finance in Claude, allows eligible US subscribers to connect selected financial accounts and ask questions about spending, upcoming bills and savings. Users can request charts, dashboards or explanations of financial activity without manually reviewing several banking applications.
According to Anthropic’s announcement, the beta is being introduced to Pro, Max and Ultra subscribers in the United States. Account connections are handled through Plaid, a financial technology company that enables applications to access information from participating banks.
Users must authorise each connection. Anthropic says Claude receives read-only access, meaning it cannot transfer money, make investments or directly manage accounts.
However, the information available to the assistant is extensive. For authorised accounts, Claude can access the financial institution’s name, account type, partial account number, balances, credit limits and up to two years of transaction history.
Those records can reveal much more than spending habits. Financial transactions may provide information about where people travel, the organisations they support and the services they use.
Anthropic says conversations within the Personal Finance section are not used to train its models. It also states that users can disconnect accounts and revoke access, while information shared outside that dedicated section follows the user’s wider privacy settings.
The company says Claude does not offer sponsored financial recommendations or promote its own loans and banking products. It acknowledges that the assistant can make mistakes and advises users to verify important figures with their financial institutions.
The development adds another category of highly personal information to AI assistants, which already help users organise documents, communications and professional tasks.
Anthropic presents the new service as a way to make financial information easier to understand. Its longer-term plans include exploring collaboration with regulated financial institutions and advisers.
3
Is AI Accelerating Its Own Development?
An independent assessment suggests Anthropic’s models are advancing roughly twice as quickly as before on well-specified tasks. Meanwhile, OpenAI has released a cybersecurity model capable of discovering serious software vulnerabilities, bringing fresh attention to the pace of AI development.
Artificial intelligence may already be accelerating the development of more capable AI systems, according to an independent assessment published on 5 November by research organisation METR.
Following 11 weeks of access to Anthropic’s internal research processes, METR estimated that the company’s progress on well-defined AI capabilities has approximately doubled relative to its previous rate. The researchers believe much of the acceleration comes from scientists and engineers using AI in their own work.
In its independent assessment, METR reported that benchmark progress had accelerated by an estimated factor of 1.9 since October 2025. The organisation cautioned that the estimate was uncertain, saying it would not be surprised if the true acceleration were closer to 1.3 or 3 times the historical rate.
The assessment examined internal data, training processes, interviews with Anthropic staff and measures of model performance. Researchers found evidence that the acceleration could not be explained simply by faster growth in the computing resources used for training.
Instead, AI tools appear to be making some of Anthropic’s researchers more productive, helping them write code, conduct experiments and improve their models.
This is different from fully autonomous AI research. METR found that AI systems were not yet making substantial independent contributions to the most important parts of Anthropic’s training process, although researchers identified signs that this could change.
The findings also come with an important transparency limitation. Anthropic exercised its right to redact a majority of the evidence METR considered most informative, citing concerns about revealing sensitive information to competing developers.
METR acknowledged that these omissions make it harder for the public to evaluate its conclusions independently.
A separate development illustrates how rapidly specialist AI capabilities are advancing.
This week, OpenAI introduced GPT-6 Cyber, a model designed to identify, validate and help repair software vulnerabilities.
The company reports that researchers used the model to develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has released fixes for many of the Firefox findings.
OpenAI classifies GPT-6 Cyber as having Critical cybersecurity capabilities under its own Preparedness Framework, a rating it says means the model can develop working zero-day exploits in many hardened real-world critical systems without human intervention. It says the model has also been trained to respect authorised testing boundaries and that access to advanced offensive capabilities is restricted to verified researchers.
The Linux and Firefox results demonstrate significant capabilities in security research. They do not establish that the model independently improves its own architecture.
Nevertheless, both developments highlight a growing tension. AI is becoming more capable at developing software, investigating vulnerabilities and assisting the researchers building future systems.
The research raises questions about whether evaluation and governance can keep pace with those improvements.
4
Can a Smartwatch Detect High Blood Pressure?
Google researchers have developed an AI system that identifies signs of hypertension using ordinary wearable sensors. The approach could make screening more accessible, although it remains exploratory and cannot replace clinical diagnosis.
A smartwatch may eventually help identify people at risk of hypertension without requiring them to use a conventional blood-pressure cuff.
Google Research announced a new AI system on 5 November that analyses information collected by sensors already found in consumer smartwatches. The system combines optical measurements of blood flow with movement data to identify patterns associated with elevated blood pressure.
Hypertension affects more than one billion people worldwide and often develops without noticeable symptoms. Many people discover they have the condition only during medical examinations or after experiencing health complications.
In the published research, scientists describe a system called AI-PPG-ACC-HTN. It processes signals from photoplethysmography sensors, which use light to measure changes associated with blood circulation, alongside accelerometer readings.
Unlike some existing cuffless blood-pressure technologies, Google’s system does not require users to calibrate it against a conventional arm cuff.
The researchers evaluated its performance against 24-hour ambulatory blood-pressure monitoring, a clinical method that repeatedly measures blood pressure throughout the day and night.
Google reports that the system achieved performance comparable to selected conventional screening approaches. In initial screening of 196 participants, it reached specificity of 90%, correctly clearing about nine in ten people without hypertension, but sensitivity of 65.8%, meaning it missed roughly a third of those with the condition.
However, these results do not mean that the system measures blood pressure with the precision of a medical device. Its purpose is to identify signs that someone may need formal assessment.
The research was conducted with collaborators from Beth Israel Deaconess Medical Center and Harvard Medical School, and the findings were published in Nature.
The potential public-health benefit is substantial. Smartwatches can collect physiological information over extended periods, potentially identifying people who might otherwise remain unaware of their condition.
There are also limitations. Smartwatch ownership is uneven, and successful screening requires access to clinicians who can confirm a diagnosis and provide treatment.
Google explicitly states that the research is exploratory and does not describe an existing or forthcoming commercial feature.
5
Can Blockchain Open Up Emerging-Market Bonds?
Schroders has brought an actively managed emerging-market debt strategy onto Solana through Centrifuge. The launch expands tokenisation beyond cash-like products, but the underlying investments retain their financial risks.
Asset manager Schroders is bringing an emerging-market bond investment strategy to the Solana blockchain, extending the range of traditional financial products available through tokenised infrastructure.
The development, announced on 4 November, involves Centrifuge, a platform specialising in bringing conventional investment assets onto blockchain networks.
According to Centrifuge’s announcement, the launch introduces tokenised exposure to the Schroder International Selection Fund Emerging Market Bond strategy, which invests in debt issued by governments and companies in emerging economies. Investors hold tokenised shares of a feeder fund, not the underlying bonds directly.
The product marks Schroders’ first deployment on Solana. The company manages more than $1 trillion across its wider investment business, although that figure does not represent the size of the tokenised product.
Until recently, much of the institutional tokenisation market concentrated on relatively simple financial instruments, particularly money-market funds, cash equivalents and short-term government debt.
Emerging-market bonds introduce additional complexity. Their performance is influenced by sovereign creditworthiness, corporate finances, currencies, interest rates and political and economic conditions in the issuing countries.
Schroders has more than 50 years of experience investing in emerging-market debt and operates a dedicated team of specialists. The company says the new arrangement will make its investment expertise available through digital financial infrastructure.
Centrifuge argues that tokenisation could allow assets managed by established institutions to connect with blockchain-based applications, lending arrangements and other financial services.
Using Solana offers access to a public blockchain rather than requiring a closed network built specifically for one financial institution.
However, issuing digital representations of an investment fund does not automatically make that investment more liquid, reduce its risks or remove regulatory restrictions on who can purchase it.
The announcement also does not establish that retail investors everywhere will be able to access the product. Eligibility, distribution arrangements and applicable financial regulations remain relevant.
The development follows a growing number of traditional asset managers experimenting with tokenised investments, suggesting that blockchain is becoming another distribution channel for established financial products.
The AI Daily Five