As AI Investing Develops Faster in 2026, Why Do You Need to Know Where the "Brakes" Are?

EmitenTrust.com SINGAPORE, Sept. 30, 2026 /PRNewswire/ -- On 26 September 2026, Singapore's Minister for Foreign Affairs, Dr Vivian Balakrishnan, delivered Singapore's national statement at the General Debate of the 81st session of the United Nations General Assembly and suggested that countries explore a UN Framework Convention on AI Safeguards. He used the image of driving a high-performance car: to go fast and stay safe, you need a powerful engine, but you equally need good brakes; and once you leave your neighbourhood and cross borders, you also need shared rules of the road, insurance and seat belts.

The analogy fits investing just as well. A global survey published by EY in April 2026 found that 49 per cent of respondents had used AI to support savings and investment decisions in the previous six months. Longbridge Academy's earlier article, AI Enabled vs AI Native: Where Brokerage AI Stands in 2026, looked at where AI sits in the investing workflow, and why the deeper AI goes, the clearer the division of roles between people and AI needs to be. This article looks at the other side of the same workflow: as AI has now gain greater ability to analyse information and take actions within investment workflows, which mechanisms make it stop when it needs to. Drawing on publications from the International Organization of Securities Commissions (IOSCO), Hong Kong's Securities and Futures Commission (SFC) and the Monetary Authority of Singapore (MAS), it sets out four problems related to these "brakes" and four tests that readers can apply for themselves.

What Are the "Brakes" on an AI Investing Tool?

The brakes on an AI investing tool are the limits on what AI is allowed to do within the investing workflow, the points at which a person steps in, and the mechanisms for stopping and correcting it when something goes wrong. They measure what AI may do and who checks it, which is a separate question from how capable its analysis is.

According to The Straits Times (republished by Asia News Network), the requirements Dr Balakrishnan set out in his speech sketch the outline of such brakes: rigorous testing and evaluation before deployment; clear limits on what autonomous systems can do; mechanisms to intervene when systems behave in ways that are not intended; and, across borders, shared rules that everyone understands. He also stressed that human beings must ultimately remain in charge. Mapped onto the AI tools investors use day to day, these requirements break down as follows:

Pre-deployment testing and evaluation are mainly the responsibility of service providers and regulators, and are hard for investors to observe directly; the other items in the table are things investors can check in everyday use.

How Far Has AI Moved Into Investment Decisions?

AI is already widely involved in the early stages of investment decisions, while the discussion of brakes among regulators and the industry is still catching up.

EY's second Global AI Sentiment Survey covered 18,152 respondents across 23 countries and territories, including Hong Kong and Singapore. It found that 49 per cent had used AI to support savings and investment decisions in the previous six months; 14 per cent had allowed AI to select financial services providers on their behalf; and over the same period, 11 per cent said they had deferred to AI to manage their finances with little or no human intervention. This is a cross-industry consumer survey and is not specific to brokerage users, but it shows that some investors are already handing certain decisions to AI.

Regulators' observations point in the same direction. In May 2026, IOSCO published its final report, Supervisory Toolkit for AI Use in Capital Markets; the survey its Fintech Task Force ran for the report received responses from 21 member authorities, including the SFC and MAS. Some respondents said the use of generative AI and AI agents in financial products and services remains at a relatively early stage, with most activity concentrated in pilots and limited-scale deployments; the survey data also shows that the speed and scale of AI development and implementation have increased over the past two years. Regulators are increasingly aware of the risks of AI agents in retail financial products and services, including control over such systems, and see a need for greater focus on governance and accountability. In November 2025, MAS published a consultation paper on proposed Guidelines on AI Risk Management, intended to apply to all financial institutions and covering generative AI and AI agents.

Dr Balakrishnan's assessment in the speech serves as a footnote to this section: given superpower rivalry and strong commercial incentives, it is too late to call for a halt to AI development; what rules do is provide assurance and confidence, so that everyone can move faster and more effectively.

What Four Problems Do AI Investing Tools Still Need to Solve?

The four problems follow the order of a single AI action: before access is granted, while the task runs, at the point of confirmation, and after data leaves a single platform.

Problem One: The Access Granted Is Often Wider Than the Task Needs

The IOSCO report describes how AI agents can have access to sensitive data, can be exposed to external content and can communicate externally; if an agent is not configured properly, the combination of these features could result in data compromise and operational and cybersecurity issues. The report suggests supervisors look at the data and systems an agent can access, the scope of actions it can take, whether those actions can be reversed, and the agent's level of autonomy, and notes that the impact of agents can be limited by designing appropriate boundaries at the planning stage. Dr Balakrishnan likewise named the loss of control over autonomous systems as one of the three main sources of AI risk.

For investors, this means a single authorisation may open up several kinds of information at once, such as quotes, holdings and cash movements. If a task that only needs quotes is also given account access, the potential impact grows if the tool is misconfigured or misused.

Problem Two: When Something Goes Wrong, It May Not Be Stopped in Time

Among the model-risk concerns listed in the IOSCO report are the absence of alert mechanisms for anomaly detection and the absence of a methodology for suspending an AI system when anomalies are detected; in the section on market risk, the report lists circuit breaker or kill switch policies among the evidence supervisors may review. The report also notes that AI agents are made up of multiple components whose interplay could create unexpected behaviour, or even cascading failures across interconnected systems, and that their complexity makes unwanted activity harder to detect. The SFC's circular on generative AI language models, issued in November 2024, similarly expects licensed corporations to have measures to promptly identify cybersecurity intrusions and, where appropriate, suspend the use of an AI language model.

These expectations are aimed mainly at firms. From an investor's point of view, stopping means something more specific: whether a task in progress can be halted, and whether access already granted can be withdrawn immediately.

Problem Three: The Confirmation Step Can Become a Formality

In the EY survey mentioned above, 14 per cent of respondents had allowed AI to select financial services providers for them, and 11 per cent had let AI manage their finances with little or no human intervention. Even where a human confirmation step exists, "automation bias" can make it hollow: the IOSCO report cautions that human overseers may tend to over-rely on AI. Among the questions it suggests supervisors ask are who can intervene or override, how quickly, and how often interventions take place. The SFC circular, for its part, expects disclosures in high-risk use cases to be made whenever a client interacts with the AI; a single disclosure at the start is not enough.

This means that whether the confirmation step works depends on what the investor sees before confirming: whether the action and its impact are clear at a glance, and whether reminders appear at every interaction.

Problem Four: Once Data Leaves a Single Platform, Who Is Responsible for What?

Dr Balakrishnan stressed that the risks of AI do not stay behind borders, which is why shared rules of the road are needed. The same issue arises within an individual investor's own set of tools: when brokerage data is connected to a general-purpose AI assistant, the data passes through at least two parties. The disclosure concerns listed in the IOSCO report include insufficient disclosure of the use of third-party AI services and over-reliance on third-party providers' disclosures. The SFC circular makes clear that its requirements apply whether an AI language model is developed by the licensed corporation itself, its group company, an external service provider or an open-source project, and that the allocation of cybersecurity responsibilities between a licensed corporation and a third-party provider should be well defined.

From an investor's point of view, three things are worth establishing before connecting a third-party tool: where the data goes, who holds the authorisation credentials, and which party is responsible for which part if something goes wrong.

What Four Tests Show Whether an AI Investing Tool Has Brakes?

Each of the four tests corresponds to one of the problems above, and they can be used to assess AI tools on any platform, including the Longbridge products described later in this article.

1. Can access be narrowed as needed and revoked at any time? Ideally, different categories of data, such as quotes, holdings and cash, can be authorised separately, and the place to revoke access is easy to find. This addresses the problem of access being wider than the task needs.
   
   
2. Is there a clear way to stop? Whether a task in progress can be halted and whether write actions can be cancelled should be answered in the official documentation. This addresses the problem of not being able to stop in time.
    
3. Can you see a full preview before confirming, and do reminders keep appearing? For actions that write, modify or involve money, the action and its impact should be clear before confirmation; reminders that you are dealing with AI and that it may be wrong should also recur during use. This addresses the problem of confirmation becoming a formality.
4. Are data flows and each party's responsibilities clearly explained? Which platforms the data passes through, who holds the credentials and which terms apply to each party should ideally be discoverable before you connect. This addresses the problem of responsibility once data leaves a single platform.

The four tests in Longbridge Academy's earlier article look at how deeply AI takes part in the investing workflow; the four tests here look at whether AI can be constrained within it. The two sets work well together.

How Do LongbridgeAI's Products Respond to These Problems?

LongbridgeAI's three products each respond to some of the four problems. LongbridgeAI comprises Chatbot (LongbridgeAI), Skill and Agent Platform, corresponding respectively to using AI directly inside the Longbridge App, connecting Longbridge's data to the AI tools investors already use, and investors designing their own investment research agents. Longbridge Group's July 2026 product launch press release describes the division of roles between people and AI across these products as a human-in-the-loop principle, with key decisions made by the investor.

LongbridgeAI's Skill: Access That Can Be Narrowed and Withdrawn

LongbridgeAI's Skill responds mainly to the first, second and fourth problems. Using OAuth authorisation, it connects Longbridge's market, fundamentals and account data to AI tools investors already use, such as ChatGPT and Claude, without the need to manage API keys manually. The security recommendations in the Longbridge MCP documentation include granting only the permissions the current task needs, starting with read-only capabilities such as market data queries, leaving credentials to be managed by the client and not copying them into untrusted environments, and revoking unused authorisations in Longbridge account security settings. Longbridge Academy's Can ChatGPT Help You Analyse Stocks? A Prompt Guide After Connecting Longbridge (2026) makes a similar point: at authorisation, you can grant market data and fundamentals access first and add account and position access later once you are sure you need it. On confirmation, the watchlist and alerts skill package in the Skill catalogue requires a second confirmation for write actions.

LongbridgeAI's Chatbot: Stating AI's Identity and Limits in the Interface

LongbridgeAI's Chatbot responds mainly to the third and fourth problems. The official notice at the foot of the LongbridgeAI product page states that LongbridgeAI is an artificial intelligence, not a real person; that it is still under development and may produce inaccurate, outdated or misleading information, which users should verify; and that its content does not constitute investment advice, with investment decisions resting on the user's own judgement. The notice also reminds users not to enter or share personal, confidential or sensitive information when interacting with it, which helps limit how much data leaves the conversation.

LongbridgeAI's Agent Platform: Setting Method and Boundaries at the Planning Stage

LongbridgeAI's Agent Platform relates to the first and fourth problems. As described in Longbridge Group's July 2026 press release, it is an open marketplace where strategies, analytical roles and skills sit side by side rather than converging on a single house view; investors can design their own investment research agent in plain language or use approaches published by others. Having investors settle the methodology an agent follows before it runs has something in common with IOSCO's idea of designing boundaries at the planning stage. On using agents published by others, Longbridge Academy's Build Your Own AI Investment Assistant: What LongbridgeAI Agent Platform Offers Hong Kong and Singapore Investors notes that it is worth understanding who published an agent before using it, and that being tested by the market is not the same as being risk-free.

Who Is Responsible When AI Gets It Wrong?

Under the current regulatory direction, licensed firms that provide AI services remain accountable for their output, while investment decisions are still made by investors, who bear the related risks.

The SFC circular states that licensed corporations remain accountable for the output of AI language models regardless of the risk mitigation measures adopted. The IOSCO report takes a consistent view: firms are responsible for the financial products and services they provide, whether or not they use AI, and the primary responsibility for addressing the hallucination risk of generative AI lies with the firm deploying the system. The report also notes that some members are considering compensation frameworks for investor losses arising from inaccurate AI outputs, with discussions still ongoing.

This means that using AI to look things up or analyse information through a service provided by a licensed firm, and doing so with a general-purpose AI tool on your own, may involve different rules and allocations of responsibility. Before using any AI tool, it is worth understanding who provides the service and whose terms apply. The availability and experience of trading-related functions in each market are subject to local regulatory requirements and Longbridge's official announcements.

How Can You Fit Brakes to the AI Tools You Use?

Four steps are a good place to start: review your authorisations, begin with read-only access, set limits in your prompts, and check and revoke access regularly. Longbridge's Chatbot, Skill and Agent Platform can all serve as a starting point, and the same practices apply to AI tools on other platforms:

1. Review your authorisations: List the AI tools currently connected to your brokerage account and check what each one has been allowed to access.
   
   
2. Begin with read-only access: For a newly connected tool, start with market data and fundamentals queries, and open up account information such as holdings only once you are sure you need it.
   
   
3. Set limits in your prompts: Say that the task is to look things up without making changes, and that any write action needs your confirmation first.
   
   
4. Check and revoke regularly: For tools you no longer use, revoke their access promptly in your account security settings.

The following prompts are designed to set boundaries for AI within a conversation and can be copied directly:

It is also advisable not to share identity document numbers, passwords or other personal details in conversations, and to keep your own judgement: AI output can serve as reference material, but investment decisions should still take your own circumstances into account.

The authorisation steps for each AI client, the scope of permissions and the security recommendations are set out on the Longbridge Skill page and in the Longbridge MCP documentation.

How Might Different Investors Use These Tests?

The test to prioritise depends mainly on how much access you have already given AI tools.

• Investors starting to use AI for research: Begin with the first test and keep access narrow, opening only market data and fundamentals queries, then consider widening it once you know the tool.

• Investors who have connected AI to their account to review holdings: Pay particular attention to the fourth test, and establish which platforms your data passes through and who holds the credentials.

• Investors considering AI agents: The second and third tests matter most; whether there is a clear way to stop, and whether you can see a full preview before confirming, determines whether "human in the loop" works in practice.

• Readers following regulatory developments: Watch for the next steps on MAS's proposed Guidelines on AI Risk Management, and IOSCO's next phase of review of emerging industry practices.

Frequently Asked Questions

What Is a UN Framework Convention on AI Safeguards?

It is an idea for an international framework that Singapore put forward for countries to explore at the UN General Assembly in September 2026. As The Straits Times explains, a framework convention, like the UN climate change convention, would be a broad international treaty that establishes foundational principles and cooperative institutions and provides the architecture for international action.

What Are the Main Parts of the "Brakes" on an AI Investing Tool?

There are four main parts: limits on access, confirmation and stopping, withdrawal and correction, and shared rules across platforms. Pre-deployment testing and evaluation are also part of the brakes, but they are mainly the responsibility of service providers and regulators.

What Is the Principle of Least Privilege for AI Access?

It means granting only the permissions the current task genuinely needs. If you only need to check quotes, for example, there is no need to open up account and holdings information as well; the narrower the access, the smaller the impact if a tool is misconfigured or misused.

What Should I Consider Before Connecting Brokerage Data to ChatGPT or Other AI Tools?

It is worth first establishing where the data goes, who holds the authorisation credentials and how responsibilities are divided if something goes wrong. In practice, you can start with read-only access and regularly check and revoke authorisations you no longer use.

Why Do AI Agents Need a Way to Be Stopped?

Because AI agents are made up of multiple components whose interplay can produce unexpected behaviour, and anomalies can be hard to detect in time. IOSCO lists the absence of anomaly alerts and suspension methods as supervisory concerns, and for investors the ability to halt a task or withdraw access at any time matters just as much.

How Do I Revoke Access I Have Given an AI Tool to My Longbridge Account?

You can review and revoke unused authorisations in your Longbridge account security settings. The Longbridge MCP documentation recommends doing this regularly and granting only the permissions the current task needs.

Final Thoughts

From the UN General Assembly to IOSCO and regulators in Hong Kong and Singapore, the focus of discussion is shifting from what AI can do to where AI stops and who checks it. Dr Balakrishnan's analogy is a reminder that good brakes are what allow a driver to speed up with confidence; for investors, a trustworthy AI tool likewise makes its access, its means of stopping, what you are confirming and where your data goes both visible and adjustable. The four tests in this article apply to AI investing tools on any platform: readers new to AI can start by narrowing access, while those considering AI agents would do well to put stopping and confirmation first.

If you already use ChatGPT, Claude or another AI tool, you can go to the Longbridge Skill page now, follow the guidance to authorise it, open only market data and fundamentals queries to begin with, and run through these four tests for yourself.

This article is for general educational and reference purposes only and does not constitute investment advice, a recommendation, or a solicitation to buy or sell any security. Any company names, tickers, funds, or other financial products mentioned or used as examples are for illustration only and do not represent an endorsement or suggestion to invest in them.

Investing involves risk, and you may lose some or all of your principal. Past performance is not indicative of future results, and the value of investments can go up as well as down. This article does not take into account your personal financial situation, objectives, or needs; please exercise independent judgement and consider seeking advice from a licensed or independent professional financial adviser before making any investment decision.

Product features, availability, and figures referenced in this article were accurate as of the stated verification date (28 September 2026) and may change; please refer to Longbridge's official channels for the most current information. This article is intended for readers in regions where Longbridge's relevant products and services are available; if the laws of your jurisdiction restrict access to this type of content, please follow your local requirements, as nothing here is intended as an offer or solicitation where that would not be permitted. Longbridge has taken reasonable care in preparing this article for reference purposes, but is unable to accept liability for any loss arising from reliance on its content.

Longbridge's AI-related tools and connectors provide access to market data and account information; they do not generate investment analysis, recommendations, or advice on Longbridge's behalf. AI-generated content may be inaccurate, incomplete, or based on outdated information, and should be independently verified before you rely on it.

View original content to download multimedia:https://www.prnewswire.com/apac/news-releases/as-ai-investing-develops-faster-in-2026-why-do-you-need-to-know-where-the-brakes-are-302894521.html

SOURCE Longbridge Group