Friday 16 October

1

Global regulators set out 12 AI safeguards for banks. But who will enforce them?

The Financial Stability Board has published new guidance for banks and financial institutions adopting artificial intelligence. The recommendations address oversight, transparency and risk, but leave a familiar question unanswered: how far can voluntary standards go?

The Financial Stability Board (FSB) has issued 12 recommended practices for financial institutions adopting artificial intelligence, calling for stronger governance, human oversight and better management of risks throughout the technology’s lifecycle.

The final guidance, published on 15 October, follows a consultation launched in June and reflects growing concern about the use of AI in financial services.

The recommendations cover how banks and other financial institutions should select, develop, deploy and monitor AI systems.

They emphasise the responsibility of boards and senior management to ensure that AI adoption is consistent with their institution’s business strategy and appetite for risk.

The FSB also calls for clear documentation of AI use, appropriate data governance, explainability, performance testing and effective human supervision. It identifies cybersecurity and dependence on third-party technology providers as areas requiring particular attention.

The guidance comes as financial institutions increasingly explore AI for fraud detection, customer service, risk assessment and internal operations.

These applications could improve efficiency and help institutions recognise patterns that conventional systems struggle to identify.

But their consequences are not always confined to operational performance.

An unreliable AI model used in a consequential financial decision could disadvantage customers, misinterpret information or introduce errors that prove difficult to identify.

More sophisticated systems may also depend on external model providers whose technologies and decision-making processes are not fully visible to the institutions using them.

That dependency creates additional questions about accountability.

A bank may remain responsible for a decision even when parts of the process rely on a model developed and maintained by another company.

The FSB recommends that institutions assess these risks from the outset and continue monitoring them after deployment.

However, the new practices are not themselves binding legislation. They are designed to be adapted to different institutions and existing regulatory frameworks.

The FSB says it will monitor developments and update its guidance as AI capabilities and applications evolve.

The approach recognises that financial institutions face different risks depending on how they use AI. Yet it also leaves regulators and individual firms considerable discretion over what constitutes adequate oversight.

Source: Financial Stability Board

2

CERN and Mistral want AI to help discover new physics. Can Europe build its own scientific models?

CERN and French AI company Mistral have agreed to explore foundation models trained on particle physics data. The partnership could advance scientific research while strengthening Europe’s capacity to develop AI on its own terms.

The European Laboratory for Particle Physics, CERN, and French artificial intelligence company Mistral have signed a statement of intent to develop advanced AI models for scientific research.

The agreement announced on 15 October establishes the basis for a potential long-term collaboration, initially focused on high-energy physics.

At the centre of the proposal is a Large Physics Model, a foundation model designed to learn from billions of simulated proton-proton collisions at CERN’s Large Hadron Collider.

The aim is to create a single system capable of supporting several tasks that currently rely on specialised algorithms.

These include reconstructing particle collisions, identifying particles, recognising unexpected signals and helping researchers evaluate future detector designs.

Particle physics presents an unusual opportunity for AI.

Experiments at the Large Hadron Collider generate extraordinary amounts of information. CERN says approximately one exabyte of experimental data has already been stored, equivalent to one million terabytes.

Analysing that information requires sophisticated computational techniques to distinguish scientifically meaningful events from the enormous volume of recorded signals.

Machine learning has been used in particle physics for years. The proposed foundation model would seek to extend that work by learning across several stages of data analysis rather than relying entirely on separate models for individual tasks.

The partnership is also intended to prepare for a future increase in data.

CERN expects the number of collisions to rise approximately tenfold when the High-Luminosity Large Hadron Collider begins operating around 2030.

An AI system capable of handling multiple analytical tasks could help researchers process increasingly complex experimental results.

However, the project remains an ambition rather than a demonstrated breakthrough.

CERN and Mistral have signed a statement of intent. They have not announced a completed Large Physics Model or evidence that it has discovered new physical phenomena.

The collaboration also has broader implications for European science and technology.

Mistral is one of Europe’s most prominent developers of large AI models, while CERN is an internationally recognised centre for fundamental research.

Combining their expertise could support scientific AI capabilities developed within Europe, potentially reducing dependence on tools controlled exclusively by American technology companies.

Yet scientific independence involves more than the nationality of the company developing a model.

Researchers also need access to training methods, adequate documentation, reproducible results and the ability to examine the system’s limitations.

Source: CERN

3

Amazon’s new AI wearables remember conversations. But who benefits from what they hear?

Amazon has introduced two personal AI devices designed to turn everyday conversations into reminders, summaries and shopping suggestions. The technology promises convenience while extending the company’s reach into a particularly sensitive source of information: what people say.

Amazon has unveiled two AI-powered devices under its Bee brand, expanding its ambitions in personal artificial intelligence with wearables designed to remember information from everyday conversations.

The company announced Bee Frontier, a device that can be worn on the wrist or clipped to clothing, and Bee Eclipse, a credit-card-sized accessory that attaches magnetically to a smartphone.

According to Amazon’s announcement on 15 October, the devices can turn spoken conversations into summaries, reminders, tasks and other information intended to help users manage their daily activities.

Bee Frontier is priced at $74.99 in the United States, while Bee Eclipse is expected to cost $99.99 when it becomes available.

Amazon acquired Bee in 2025 and is now expanding the technology’s integration with its existing services, including Alexa+ and Amazon.com.

The devices are designed to capture information that might otherwise be forgotten.

Someone discussing plans with a colleague could receive a reminder about an agreed meeting. A conversation about a task could become an item on a to-do list.

Users can also connect Bee to services such as Gmail and Google Calendar, allowing information gathered from conversations to be incorporated into other aspects of their digital lives.

But Amazon’s plans extend beyond personal organisation.

The company is introducing connections between Bee and its retail platform, allowing information from conversations to inform shopping suggestions.

A person mentioning a household product they need could subsequently receive a relevant recommendation through Amazon’s services.

This creates a potentially valuable commercial opportunity, but it also changes the relationship between personal assistance and advertising.

A device designed to help people remember what they say may also generate information about their interests, intentions and purchasing needs.

Amazon says audio is processed in real time and deleted rather than retained as recordings. Conversation transcripts are encrypted, with access controlled by the user.

The company also says the devices include visible recording indicators and controls intended to protect privacy.

These are important safeguards. However, they do not eliminate questions about whether other people present during a conversation understand that information is being processed.

Consent becomes particularly complicated when wearable technology operates in workplaces, social gatherings or shared public environments.

There is also the question of how personal information may influence commercial recommendations.

The value of a digital assistant lies partly in its ability to understand context. But the more context it collects, the more important it becomes to distinguish assistance requested by the user from suggestions shaped by the platform’s commercial interests.

Amazon is entering an increasingly competitive market for personal AI devices, where companies are experimenting with alternatives to smartphones and conventional voice assistants.

The challenge will be demonstrating that the benefits outweigh the practical and privacy costs of incorporating AI into ordinary conversations.

Source: Amazon

4

Nvidia launches an open AI model for complex tasks. Does customisable AI challenge the biggest players?

Nvidia has introduced Nemotron 3.5 Super VL, a multimodal model designed for specialised AI agents. Its emphasis on efficiency and customisation could give organisations more flexibility, although access to computing power remains a significant constraint.

Nvidia has introduced Nemotron 3.5 Super VL, an open-weight artificial intelligence model capable of analysing text, images, documents and video while performing complex tasks involving reasoning and software tools.

The model, announced on 15 October, is designed for organisations developing specialised AI agents rather than relying entirely on general-purpose commercial models.

It uses a mixture-of-experts architecture containing 120 billion parameters, but activates approximately 12 billion parameters for each generated token.

This design aims to reduce the computational resources needed to operate the model while retaining substantial capacity for complex tasks.

Nvidia says Nemotron 3.5 Super VL can deliver up to five times greater throughput than certain comparable open models in its size category under specified test conditions.

The model combines reasoning, instruction following and tool use with the ability to interpret several types of information.

Potential applications include analysing financial documents, identifying software vulnerabilities, extracting information from charts and reviewing video.

Organisations can also customise the model using their own data, policies and operational requirements.

Nvidia has released model checkpoints, training resources and supporting datasets, allowing developers to adapt the technology rather than depending exclusively on a remotely hosted proprietary model.

This could be particularly useful for businesses that require systems tailored to specialised workflows.

For example, a financial services company might adapt an AI model to process internal documents according to its own procedures. A cybersecurity team could train a system to recognise particular patterns of threats.

Nvidia has cited examples of companies reporting improvements after customising Nemotron for specific tasks.

These include lower inference costs, reduced processing times and improved accuracy in selected workflows.

However, such results should be understood within their testing conditions. Performance on one task does not establish that a model will outperform alternatives across every application.

There is also a distinction between an open-weight model and a fully open-source AI system.

Releasing model weights and training resources allows greater inspection and customisation, but does not necessarily provide unrestricted access to all training data, development decisions or underlying infrastructure.

Nor does it remove the cost of operating advanced AI.

A 120-billion-parameter model still requires substantial computing resources, particularly when deployed to serve many users or process complex material.

Nvidia’s position is therefore interesting.

The company sells much of the hardware underpinning the AI industry, while increasingly developing models and software intended to support broader adoption.

Offering customisable AI could help organisations reduce dependence on proprietary model providers while reinforcing demand for the infrastructure needed to operate those models.

Source: Nvidia

5

GitHub introduces AI tools to find and fix security flaws. But who checks the AI?

GitHub has made two AI-assisted security tools generally available, helping developers identify vulnerabilities and propose repairs. The development could improve software security while raising new questions about how automatically generated fixes should be verified.

GitHub has introduced two generally available AI-assisted security tools designed to identify vulnerabilities in software and help developers repair them.

The features, AI Scan for pull requests and Agentic Autofix for code scanning alerts, were announced on 15 October as part of the company’s expanding security capabilities.

AI Scan examines changes submitted through pull requests, the process developers use to propose and review modifications to software.

It complements GitHub’s existing CodeQL analysis by identifying potential vulnerabilities in languages and frameworks that conventional scanning tools may not fully cover.

Agentic Autofix takes the process further.

Rather than simply identifying a security problem, the tool can examine relevant files across a codebase and propose changes intended to address the vulnerability.

Both developments reflect a broader shift in software engineering.

Generative AI is already being used to write code, explain unfamiliar systems and automate repetitive development tasks.

Now it is increasingly being applied to evaluating the security of that same software.

The potential benefits are considerable.

Security weaknesses can be difficult to detect, particularly in large applications containing extensive collections of dependencies and interacting components.

AI-assisted scanning could help developers identify problems earlier, while automated suggestions may reduce the time required to investigate and repair vulnerabilities.

For smaller development teams, such tools could also make sophisticated security practices more accessible.

However, identifying a possible vulnerability and fixing it correctly are different problems.

A proposed repair may address one security weakness while introducing another defect, changing expected software behaviour or failing to account for how different components interact.

That is why human review and conventional testing remain essential.

There is also a question about the evidence used to establish the effectiveness of AI-assisted security tools.

Developers need to understand how frequently a system identifies genuine problems, how often it produces false alarms and whether its proposed fixes remain effective under realistic conditions.

These questions become particularly important as organisations adopt AI-generated code at greater scale.

If one AI system writes software and another AI system evaluates its security, the development process may become faster. But it also risks creating an impression of independent verification where both systems have limitations that developers do not fully understand.

GitHub’s new tools are intended to support developers rather than eliminate their responsibility for reviewing changes.

The company will need to demonstrate that improved automation translates into fewer exploitable vulnerabilities in deployed software, not merely larger numbers of alerts and suggested fixes.

Source: GitHub