Last Saturday, Singapore stood up at the UN General Assembly and asked the world for something I have been harping on about for a while now. Put the AI safeguards in place first, before things go out of whack.
Foreign Minister Dr Vivian Balakrishnan was delivering Singapore’s national statement in New York, and AI took up a fair chunk of his 20 minutes on stage. His view was that AI’s capability is “advancing faster than our appreciation of its risk.” For a speech at the UN, that is about as blunt as diplomacy gets.
What he actually said
The headline idea was a UN Framework Convention on AI Safeguards, along with a possible international body that could do for AI what existing institutions already do for common technical standards in other fields.
He then got quite specific. Countries should have comparable ways of testing AI, and serious incidents should be reported quickly across borders, even where governments disagree on how to regulate. He also wanted testing before deployment, limits on autonomous systems, and ways to step in when those systems behave unexpectedly.
And in a very Singaporean move, he offered to share our own governance frameworks and testing tools with other countries. Sharing is caring, as they say.
To me, it sounded so practical. With the US and China racing ahead, he acknowledged it is too late to ask for a safety pause. What would help, he argued, are institutions that can set the standards and safeguards everyone works to. The full speech is below for reference.
And it is not only governments saying this. Earlier this month, the frontier AI bosses themselves lined up behind Dario Amodei’s call to pace the frontier, with Sam Altman, Elon Musk and Demis Hassabis all agreeing within days. Mind you, Anthropic then released Claude Opus 5.5 barely two weeks later, its first model since Amodei made that call. To be fair, it performs at the level of their existing top model rather than beyond it, so the frontier itself has not moved. Still, the timing did raise an eyebrow or two.
Why this sounded so familiar to me
When I wrote about the AI industry asking itself to tap the brake two weeks ago, I compared the race to a car doing 200 kilometres an hour on a road built for 80. So I smirked a little when Minister Balakrishnan reached for cars too. A high-performance car, he said, needs a powerful engine and equally good brakes, and once you start crossing borders you need traffic lights and seat belts as well. Traffic lights have never stopped anyone from driving. If anything, they are the reason the rest of us get home in one piece.
This belief goes back well before AI for me. My years looking after corporate brand at Disney taught me that protecting a brand’s integrity and its intellectual property is far easier upfront than trying to repair it afterwards. Once something slips out into the world, clawing it back is a whole other job.
Three ideas from his speech translate neatly into how any organisation should be using AI today, and each one makes a practical case for AI safeguards at any scale.
Why AI safeguards can’t wait for the law
It helps to draw a clear line between three things that often get lumped together, legal compliance, technical soundness and ethical responsibility. A system can tick every regulatory box and work exactly as designed, and still fail because people stop trusting it, or because it causes harm nobody saw coming.
Ethics carries the most weight where regulation is still catching up, and that is exactly the gap Minister Balakrishnan was describing. If capability is outrunning our understanding of risk, then waiting for the law before putting AI safeguards in place simply means waiting too long. My advice is to decide what responsible use looks like for your team before a regulator decides it for you.
Break it before someone else does
Red teaming is one of the most useful habits any organisation can pick up. It means deliberately trying to break a system before it launches, so you find the problems before the public or the regulators find them for you. For most businesses, that can be as simple as asking a colleague to try and trip up your customer chatbot before it goes live.
His call for testing before deployment is that same instinct, only this time at the level of nations. It also echoes the OpenAI sandbox breach I wrote about in July. An independent red team had already caught the model trying to cheat its own evaluations, and the test went ahead anyway with its AI safeguards switched off. But wouldn’t you know it, the model found its way out.
Testing only earns its keep when the findings change what you do next. So Singapore offering its testing tools to other countries is a lovely touch. Think of it as red teaming, shared as a public good.
The real work starts after launch
This is the one I feel most strongly about. Of every stage in an AI system’s life, monitoring after it goes live is the one most often neglected. Once a system goes live, attention tends to drift to the next shiny project, which is precisely when the data shifts and the unintended effects start showing up at scale.
That same sandbox breach makes the case rather well. Hugging Face, an outside company, detected and contained the intrusion first, and OpenAI only connected it to its own test five days later. That five-day gap is exactly what cross-border incident reporting is meant to close.
It is also why his call for ways to intervene when autonomous systems misbehave deserves a proper hearing. AI safeguards that stop at launch day are only doing half the job, so put someone in charge of checking outputs regularly, and agree upfront what happens when something looks off.
So what does this mean for the rest of us
There is something rather fitting about Singapore making this case. Small states rarely get to shape technology on their own terms, so shared rules and common AI safeguards tend to protect the smaller players most of all. As a Singaporean, I felt a little pride watching that argument land on the world stage.
None of this came out of nowhere either. Singapore published its first National AI Strategy back in 2019, the same year it launched its Model AI Governance Framework, one of the first of its kind anywhere. In 2022 it followed up with AI Verify, billed as the world’s first AI governance testing toolkit, and a refreshed national strategy arrived in December 2023 under the banner of AI for the public good, for Singapore and the world. So when Minister Balakrishnan offered to share Singapore’s testing tools, he was offering something the country has spent years building.
The newest piece landed just ten days before the speech, when Prime Minister Lawrence Wong launched a curated catalogue of more than 200 AI courses at the SkillsFuture Festival, and eligible Singaporeans who attend one can now claim six months of a premium AI tool for free, from ChatGPT Plus to Google AI Pro. It is a message the government repeats at almost every opportunity, and Wong put it plainly at the launch, “Whatever jobs we have, they will not be the same as today.” Safeguards abroad and skills at home are two halves of the same idea.
The same logic holds inside a business. Most of us in Southeast Asia are using AI tools built elsewhere rather than building our own, and a vendor’s guardrails will only take you so far. We do not need to wait for a UN convention before building AI safeguards of our own. Know where legal ends and trust begins, test before you launch, and keep watching after you do.
So where does your organisation sit right now? Are the safeguards being built alongside the tools, or somewhere after them?
