Day: October 15, 2026

Singapore’s war on scams has mostly been fought far from the victim: in police operations centres, bank back offices and, more recently, in rulebooks telling platforms what they must block. Now the city-state wants part of that fight to happen inside the phone itself.

Google on Wednesday announced two initiatives with the Singapore Government, unveiled on the sidelines of Singapore International Cyber Week (SICW) 2026. The first brings Singapore into Google’s Fairwind Programme, a gated scheme that gives vetted defenders the company’s most capable cyber models, including its new frontier model, Gemini 4 Argon.

The second pairs GovTech Singapore with Google DeepMind to build an on-device AI model for Android that flags scam attempts in real time without sending private conversations to the cloud.

Also Read: Singapore’s cybersecurity paradox: Leading in digital, lagging in defense

Both deals are framed as defensive. Both also deepen Singapore’s reliance on a single US tech giant at a time when access to frontier AI has become a geopolitical lever.

An unguarded model for trusted hands

Fairwind launched on September 2 as a limited access programme for governments and trusted partners, starting with Gemini 3.8 Flash Cyber paired with Google’s CodeMender agent. Argon, announced on September 30, has since become the programme’s headline act. Fairwind partners and internal Google teams get a version of Argon without its cyber guardrails, the filters that normally block requests that look like hacking.

Organisations may only grant Argon access to internal cybersecurity, incident response, or penetration testing teams, and must track employee access and use.

Google’s pitch is speed. Instead of spending weeks manually fixing vulnerabilities, it says defenders can generate verified, deployment-ready patches in minutes inside their own secure cloud environment. On Google’s own leaderboard, Grok 4.7, Gemini 4 Argon and GPT-6 Astra tie at 68 per cent on the CWE-bench v1 remediation benchmark, which also means the model still misses roughly a third of the test’s tasks.

“Minutes, not weeks” is a vendor claim, not a measured outcome. Analysts have pointed out that the clearest evidence would be merged-patch or time-to-remediation results from Fairwind partners, which Google has not published. An AI that finds bugs faster than humans can review fixes risks lengthening the backlog, not shortening it.

There is also the question of who sees what. Google says it monitors Argon’s chain of thought and its actions and stops execution when necessary. That is a sensible safety measure, though it implies a foreign company has some line of sight into how a government’s defenders probe their own systems. Google does say Argon supports zero data retention when used as a managed model on Gemini Enterprise, but the terms of Singapore’s participation were not disclosed.

A crowded, fragile club

Google is not the only lab rationing its sharpest cyber tools. Anthropic launched Project Glasswing on April 7, 2026, deploying its restricted Claude Mythos Preview model to a vetted consortium, and OpenAI responded with Daybreak on May 11. Each programme decides which governments and companies qualify, making access itself the product.

That access can disappear overnight. In June, US export controls forced Anthropic to suspend its Mythos and Fable models before Washington lifted the restrictions at the end of the month, an episode e27 has argued created a hydra rather than containing risk. For a small state building critical defences on American models, the lesson is clear: the off switch sits in another capital.

The scam problem is a human problem

The second initiative targets a more personal threat. Singapore’s scam-related losses fell 19 per cent to S$913.1 million (about US$715 million) in 2025, but the composition of those losses explains why the government is looking at devices. In 81.8 per cent of 2025 scam cases, victims transferred the money themselves rather than having their accounts taken over. Takedowns and bank freezes mostly come too late; the decisive moment is when a manipulated person presses “send”.

Also Read: Deepfake fraud losses hit US$3.7B as scams spread beyond social media

The GovTech-DeepMind model is designed for that moment. Lightweight Gemma models are fine-tuned on local scam narratives, including impersonation of public institutions, banks and local figures, and run entirely on the handset. When a conversation shows the high-pressure patterns of social engineering, a warning pops up on screen.

“Fighting scams shouldn’t mean sacrificing personal privacy,” said Mark Chen, Director of Trust and Safety at GovTech Singapore, adding that the approach keeps personal communications completely private.

The announcement leaves key questions unanswered. There is no launch date, no accuracy or false-positive data, and no explanation of how the model reads conversations inside end-to-end encrypted apps like WhatsApp or Telegram, which by design hide content from everyone but the sender and recipient. The model is also Android-only in a market where iPhones are hugely popular, leaving a large share of users uncovered.

The timing is not accidental. Singapore has tightened scam rules for messaging, social media and e-commerce platforms, with Google Messages and Google Meet among the designated services that must comply by January 31, 2027, under threat of fines of up to S$10 million (US$7.83 million). Meta has already disrupted 3.7 million scam-linked assets alongside police, and authorities have taken down 30,000 iMessage accounts. A government-endorsed on-device shield is as much Google’s compliance story as its public-safety one.

It also enters a crowded field. Singapore’s ScamShield app already filters calls and messages, and caller-ID players such as Taiwan’s Gogolook (Whoscall) and Truecaller have built large regional followings. Google’s edge is distribution: a model built into Android reaches people who would never download a separate app.

Grading its own homework

Google surrounded the announcement with supporting statistics. It says Google Play Protect has blocked more than 3.3 million installation attempts of potentially malicious apps across more than 770,000 devices, without specifying the period. It also cites studies showing that ten minutes of its Be Scam Ready game reduced players’ estimated scam risk by 20 per cent, and is launching a 2.0 version that halves playtime. These figures come from Google, and the effects of a quiz on behaviour months later are hard to measure. Its earlier work with the region, such as the US$5 million ASEAN Foundation drive, followed a similar mix of philanthropy and product.

Also Read: Should cybersecurity be nationalised?

If the on-device model works, the stakes reach well beyond Singapore. Malaysia ranked highest in the region for scam prevalence, with 32 per cent of adults losing money to scams, followed by the Philippines at 31 per cent. In those markets, as e27 has argued on Thailand’s scam epidemic, the problem is now technological and so must be the response, especially as AI phishing turns trust into the weakest link.

Singapore has once again cast itself as the region’s test bed. Whether this test produces a public shield or a deeper dependency will depend on details neither side has shared yet.

The post Singapore wants Google’s AI to catch scammers before victims hit send appeared first on e27.

In early 2018, Yujun Chean and Prateek Jogani were selling car insurance with a cashback twist. Their Singapore startup, Vouch Insurtech, had just raised SGD1 million (then US$755,000) in seed funding as the parent of Thai insurtech FairDee. Eight years on, the pair are back with a much bigger cheque and a far more crowded market.

Their new company, Parallel Loop, has launched with US$6.6 million in seed funding led by Bengaluru-based 3one4 Capital. Picus Capital, Flourish Ventures and Economic Development Innovations Singapore (EDIS) also joined, alongside three high-profile angels: Paytm founder and CEO Vijay Shekhar Sharma, CRED founder Kunal Shah, and Philip Yeo, former chairman of Singapore’s Economic Development Board and A*STAR.

Also Read: Everyone is talking about AI adoption. But are enterprises actually ready for it?

Founded only in June 2026, Parallel Loop calls itself an “AI operating partner” for mid-market and large enterprises. Four months in, it has 30 staff across Singapore, Bengaluru, Bangkok, London and Dubai, and more than 10 clients. These include Indian meat-delivery firm Licious, electricals maker Havells and Tokio Marine Safety Insurance (Thailand). Chean told e27 that Singapore enterprises, including a local healthcare group, are also on the list.

“Our aim is to build a global AI company,” he said.

The gap it is chasing

The pitch rests on a problem most enterprise leaders will recognise. AI tools have made individual employees faster, but turning that into a company-wide capability is far harder. The rules, definitions and know-how that make a business tick are scattered across systems, meetings, inboxes and chat apps. A model that cannot see that context is a clever intern who skipped onboarding.

“Intelligence has become abundant. What many organisations lack is the context, systems and capability to turn this into tangible business outcomes,” said Jogani.

The diagnosis fits the region. Southeast Asia, the argument increasingly goes, does not have an AI adoption problem so much as a scaling problem, and most enterprise AI in APAC is still stuck in the proof-of-concept room.

Three products, one bet

Parallel Loop’s offering has three parts:

• AI Enablement works with leadership teams to identify high-value use cases and build roadmaps.
• The AI Stack is a modular platform that connects a client’s existing systems with its organisational context. It powers everything from live dashboards to deterministic workflows, meaning automations that produce the same result every time rather than the varied output typical of generative models.
• The Software Factory pairs forward-deployed engineers with AI agents to modernise legacy systems and build new applications. Those forward-deployed engineers are the centre of gravity. Popularised by Palantir, the model embeds a vendor’s engineers inside client teams instead of handing over software and walking away.

The stack runs in the client’s own cloud or on premises, with data residency set by the client. It works with multiple model providers, including open-weight models, and keeps an audit trail of prompts, responses, context changes and actions. Clients retain ownership of their data and of the work built for them. That speaks directly to the regional appetite for sovereign AI, where, as one recent analysis argued, sovereignty starts long before the model itself.

Also Read: Why Southeast Asia cannot build sovereign AI on borrowed choices

Licious is the showcase client. Co-founder Abhay Hanjura said the company set out this year to become “fully AI-native”. He said teams across every function now use Parallel Loop daily as “a shared intelligence layer across the organisation”.

A familiar founding duo

The founders’ track record goes a long way towards explaining why investors moved so quickly. Jogani and Chean have worked together for nearly a decade. After Qoala acquired FairDee in 2021, Jogani became Group CTO and Chean Group COO of the Indonesian insurtech.

Qoala grew into five markets, processing millions of policies a month, and raised more than US$130 million from investors, including a US$65 million Series B in 2022. Jogani, who brings around two decades of technology experience, began running hands-on AI programmes for leadership teams in 2025. He noticed that many teams were adopting AI tools, but few were building the systems and processes needed to sustain the gains.

“They’re closing enterprise deals and expanding accounts across multiple geographies with an outcome-led approach,” said Anurag Ramdasan, Partner at 3one4 Capital.

A crowded field

Parallel Loop is walking into a brawl. Global consultancies such as Accenture and Deloitte sell AI transformation programmes to the same boardrooms. AI labs, including OpenAI, have built forward-deployed engineering teams of their own to win enterprise accounts, and Palantir still sets the template. Closer to home, Singtel’s RE:AI has partnered the SIT-NVIDIA centre to tackle the enterprise AI deployment gap.

The model also carries a familiar risk. Forward-deployed engineers are expensive, and the more a startup leans on them, the more it resembles a consultancy: software margins on the pitch deck, services margins on the books. Thirty people serving more than 10 clients across five cities in four months is fast. Whether the AI Stack becomes a reusable product, or each engagement turns bespoke, will decide whether Parallel Loop is valued as a software company or a services firm.

Also Read: Singapore firms embrace agentic AI, but audit trails remain thin

There is a quieter tension, too. A stack that promises no model lock-in also makes it easier for clients to walk away once the capability is built in-house.

For now, the founders have the network, the cheque and a client list that spans three countries. The harder test is whether enterprises that have been told to stop automating broken processes and calling it AI transformation will pay someone to fix the process first.

The post Parallel Loop nets US$6.6M to sell enterprises the one thing AI models lack: context appeared first on e27.

An AI chatbot that drafts a bad email is an embarrassment. An AI agent that sends it, to the wrong customer, with the wrong file attached, after approving the wrong refund, is a liability. Singapore’s regulators have clearly been losing sleep over the second scenario.

On Wednesday, the Infocomm Media Development Authority (IMDA) announced three initiatives at the AI Assurance Exchange: The Agentic Edition, an event organised by the AI Verify Foundation (AIVF) and SGTech that drew more than 300 deployers and testers. They are a starter kit for testing agentic AI systems, an accreditation programme for third-party AI testers with five firms in its first batch, and a new taskforce to defend the infocomm and media sector against AI-enabled cyber threats.

Also Read: What AI safety researchers actually worry about

Individually, each is a modest piece of guidance. Together, they signal something bigger: Singapore is moving from telling companies how AI governance should work to building an industry whose job is to prove that it does.

Why agents need a different exam

Traditional software follows rules. Generative AI produces content. Agentic AI does something more unnerving: it reasons, acts on a user’s behalf, touches sensitive data and changes the systems around it. Testing only the final output, as most companies do with chatbots, misses most of what can go wrong.

The Starter Kit for Testing Agentic Systems for Safety and Reliability tries to fill that gap with a four-step approach: identify risks such as erroneous or unauthorised actions, design realistic tests without causing real-world damage, run the tests and analyse the logs, then monitor and retest continuously.

The more interesting part is the new vocabulary. The kit asks testers to examine an agent’s entire “trajectory”, meaning both its reasoning and its actions, and to track agent-specific metrics. Tool use accuracy measures whether the agent called the right tool. Handoff accuracy checks whether it passed the task to the right agent or human. Human-in-the-loop rates show how often a person had to step in. It also tackles a problem most enterprises will recognise: how to test a system when you have only black-box visibility, because you bought the agent from a vendor rather than built it.

The kit draws on consultations with government agencies and more than 20 deployers and testers, including Google, AWS and Resaro, and includes case studies from Cohere, CyberSierra, Earlybird AI, Thailand’s KBTG and Knovel Engineering. It extends Singapore’s Model AI Governance Framework for Agentic AI, released earlier this year, and an earlier starter kit for testing large language model apps. IMDA calls it a living document, which is a polite way of admitting that nobody has fully worked this out yet.

The demand is real. In August, a Sumsub and Singapore Fintech Association study found that 94 per cent of Singapore businesses were using or piloting multi-step AI systems, yet only 29 per cent could produce an audit trail for AI-driven decisions. Ninety-eight per cent said they were ready to adopt third-party verification. The market for checking agents exists. The question was who would do the checking.

A badge for the testers

AIVF’s answer is the AI Tester Accreditation Programme (AI TAP), which it bills as a first-of-its-kind scheme for third-party AI testing providers. Applicants go through technical, operational and financial evaluations covering testing competency, track record, delivery capability and financial sustainability. That last criterion is telling. Regulators clearly don’t want an assurance firm folding halfway through an engagement.

Also Read: From copilots to colleagues: How agentic AI is redefining enterprise productivity

The first five accredited firms are AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan. Accreditation lasts one year, so firms have to keep earning it as the technology changes.

The stakes go well beyond enterprise IT. AIVF says the testers will assess applications in healthcare, finance, hiring and public services, the areas where an unchecked agent can quietly reject a job applicant, deny a claim or misroute a citizen’s request without anyone noticing.

Vulcan co-founder Alex Leung put the case bluntly: when AI can “send emails, move data or approve transactions, testing isn’t optional.” Resaro co-CEO April Chin argued that organisations need evidence from testing done outside the development team, not just from the people who built the system.

The questions the badge doesn’t answer

The programme deserves some scrutiny. The release does not say whether any sector regulator will require accredited testing. Without that, AI TAP risks becoming a marketing badge rather than a safeguard.

There is also a small-world problem. Resaro helped inform the starter kit, and Knovel Engineering appears in its case studies. Both are now accredited testers. In a market this young, that overlap is probably unavoidable and not improper. But financial auditing offers a clear lesson about what happens when the firms that write the standards, sell the advice and certify the results are the same few names. PwC’s presence in the first cohort makes the comparison hard to ignore.

Cost matters too. Large banks can afford independent AI audits. A Series A startup in Jakarta or Ho Chi Minh City running customer-facing agents may not, which leaves the free starter kit as its realistic option. Singapore has long pitched its governance toolkits as templates for ASEAN neighbours, and Malaysia is preparing an AI Governance Bill. Whether AI TAP travels across the region will depend on whether it stays affordable.

A taskforce on a 2027 deadline

The third initiative, the Frontier AI Resilience and Security Taskforce (FAIRS), brings together more than 20 leaders from telecommunications, media, cybersecurity and AI. Its focus is the risk that frontier models are getting better at finding vulnerabilities in networks that businesses and government systems depend on.

It has three priorities: faster response to AI-accelerated threats, segmenting critical networks to stop disruption from spreading, and pooling sector-wide intelligence on new attack techniques. Its first report is due in the first half of 2027. Attackers armed with the same models, already behind a wave of deepfake fraud, are unlikely to wait that long.

Also Read: Nearly half of APAC enterprises commit US$1M+ to agentic AI

Still, the direction is right. As e27 has argued, AI governance is moving from promises to proof. Singapore has now built the exam, chosen the first examiners and set a renewal date. Whether anyone is required to sit the test is the next policy fight.

The post Singapore’s next AI industry may be the people who check the agents appeared first on e27.

Anthropic has finally given its Singapore customers somewhere to knock. The San Francisco-based AI lab opened an office in the city-state on Wednesday, its fifth in the Asia Pacific, and signed a memorandum of understanding (MOU) with the Ministry of Digital Development and Information (MDDI) that puts AI safety testing at the centre of the relationship.

The MOU was signed on Tuesday, on the sidelines of Singapore International Cyber Week, by MDDI Permanent Secretary Chng Kai Fong and Anthropic co-founder Ben Mann, with Minister for Digital Development and Information Josephine Teo and Anthropic’s Head of APAC Policy Michael Sellitto looking on. It covers three areas: AI safety, security and model evaluation; research on how AI is adopted and what it does to the economy; and responsible adoption across government and industry.

Also Read: Singapore wants Google’s AI to catch scammers before victims hit send

The timing is hardly accidental. The same week, the Infocomm Media Development Authority (IMDA) unveiled new assurance measures for AI agents, a sign that Singapore’s regulators now worry more about software that acts than software that chats.

What is actually in the deal

Strip away the diplomatic language, and the MOU commits Anthropic to technical exchanges with Singapore’s AI safety, security and assurance institutions on model evaluation, testing standards and incident response, the playbook for what happens when a model misbehaves in the wild. Anthropic also says it will share research and data from its Anthropic Economic Index, help small and medium-sized enterprises (SMEs) adopt AI, build AI skills among public officers, and offer policy input as Singapore drafts its AI rules.

None of this is new territory. Anthropic joined the Generative AI Evaluation Sandbox run by IMDA and the AI Verify Foundation in 2023, and later worked with them on red-teaming (deliberately attacking a model to expose its weaknesses) in English, Mandarin, Malay and Tamil. The MOU formalises a relationship that already existed.

What the announcement leaves out is just as telling. There is no investment figure, no headcount target for the office beyond a note that it is hiring, and no word on whether the SME and skills programmes come with money attached. MOUs are statements of intent, not contracts, and Singapore signs plenty of them.

The customer list is the real story

Anthropic’s strongest case for an office here is the business it already has. More than 1,500 engineers and technical staff at the Government Technology Agency of Singapore (GovTech Singapore) use Claude Code, its AI coding tool, to modernise government systems. The Cyber Security Agency of Singapore is exploring how Claude could support the country’s cyber defence.

In finance, DBS uses Claude for software development and productivity in its Singapore operations, while sovereign wealth fund GIC is rolling it out across investment and corporate services teams. Regulated firms can now run Claude with in-region inference on Amazon Bedrock, meaning their data is processed within the region. That matters now that the Monetary Authority of Singapore has given financial firms a year to prepare for AI risk rules.

Also Read: Why Southeast Asia cannot build sovereign AI on borrowed choices

The consumer internet giants are in too. Grab says BriX, its Claude-powered sales assistant, saves its sales teams around 40,000 hours a quarter. Sea uses Claude to make Shopee’s customer service chatbot faster and more accurate, and thousands of its employees write and debug code with it. Supabase, the developer platform with Singaporean roots, is building with Claude Code.

A caveat: these figures come from Anthropic and its customers, not independent audits. So does the claim that Singapore ranks second of 121 countries for Claude.ai usage relative to population, at nearly six times what its size would predict. Usage per head also says little about value created, a gap visible in research showing messy data keeps many SME pilots stuck.

“Our customers here hold AI to an extremely high standard, especially when it comes to reliability and security,” said Dale Finlay, Anthropic’s General Manager for ASEAN.

Rivals are circling the same prize

Anthropic is not alone. OpenAI last month hired former Elastic and Snowflake executive Sanjay Deshmukh to lead Asia Pacific sales, and has backed a public-private AI accelerator in Thailand. Google is betting on schools, telcos and local languages to scale Gemini across the region. Databricks has unveiled a US$350 million investment plan for the city-state. And Singapore’s own SEA-LION model offers agencies a homegrown option for Southeast Asian languages.

What sets Anthropic’s pitch apart is that it leads with safety evaluation rather than consumer reach, the area where Singapore, through AI Verify, has made itself a global reference point. For a lab whose brand rests on pairing capability with safety, a government that tests models for a living is a natural partner.

The questions an MOU cannot answer

Two uncomfortable issues sit beneath the handshake photos.

The first is dependency. In June, Anthropic suspended access to its Fable and Mythos models for nearly three weeks to comply with US Department of Commerce export controls, restoring it on 1 July. That episode was a reminder that Washington, not Singapore, holds the off switch for America’s frontier AI. The MOU says nothing about that risk, and it is fair to ask how GovTech would cope if a similar order hit the models its engineers now rely on.

The second is commercial. Anthropic’s IPO prospectus shows two customers generated nearly a quarter of its 2025 revenue, while the company plans to spend US$518 billion on compute and infrastructure. Spreading revenue beyond a few big US clients is a necessity, and a market of wealthy, AI-hungry institutions is where that starts. The office is a safety partnership, but it is also a sales operation.

Also Read: Singapore’s AI dividend will depend on what happens after the pilot phase

Neither point makes this a bad deal for Singapore. Closer access to a frontier lab’s evaluation methods strengthens its claim to be the region’s AI assurance hub, and local developers, more than 300 of whom attended Anthropic’s Builder Summit this week, gain a shorter line to the company.

The real test will come later, in what the MOU produces: published testing standards, incident-response protocols that get used, and SME programmes with funding behind them. Until then, Singapore has a new tenant with an impressive customer list, and another piece of paper.

The post Anthropic’s Singapore MOU bets on safety testing, not consumer reach appeared first on e27.

Anyone who has spent 10 minutes pressing “0” to escape a voice bot knows how the last automation wave ended. Call centres were promised dramatic savings from IVRs and chatbots. Customers got phone trees that went in circles, and companies got a fresh layer of complaints to manage.

Sunny Rao sold that promise from the inside. Before becoming SVP of APAC at JFrog, he spent years at Avaya, Genesys and Nuance, three firms that defined the contact-centre automation era. Now he is watching Southeast Asia fall for AI agents, and he thinks the region risks repeating an old mistake: confusing a faster task with a more productive business.

Also Read: Singaporean founders’ Lightsage bags US$4M to decode how AI agents choose software

“An agent completing a task faster does not necessarily translate into greater productivity, particularly if it creates additional review, remediation or security work downstream,” he tells e27 in an email interview.

The adoption-readiness gap

The enthusiasm is not in doubt. A recent Agoda AI Developer Report found that 53 per cent of developers in the region say agents are already in production or broad use. Only 38 per cent believe their codebase is ready for that autonomy, a readiness gap e27 has examined before, and one that explains why regional tech leaders are learning to trust agents, but not with production.

Rao, who speaks regularly with CIOs in Singapore, Jakarta and Bangkok, frames the problem as one of control rather than capability. Coding agents now select dependencies, pull tools, modify code and interact with enterprise systems at machine speed. A company may know an agent is running without knowing what it consumed, which permissions it used or whether its actions complied with policy.

“The real readiness test is not whether an enterprise can deploy autonomous agents, but whether it can detect and control unintended actions before they become business risks,” he says. The worry has only sharpened since OpenAI shelved GPT-6.1 Astra, a model built to handle complex tasks with less human supervision.

Tokens are the wrong yardstick

Most companies still measure AI spend in seats or tokens. Rao thinks both miss the point. If a CFO could track one number, he would pick cost per successful, production-ready outcome.

The logic is simple arithmetic. An agent that is cheap to run but needs repeated prompting, produces unusable work or creates hours of human clean-up may cost more than a pricier one that gets it right first time. Tokens and licences are inputs, not value, a point that echoes e27‘s look at the hidden economics of autonomous AI agents.

Also Read: Agentic AI ambitions in Singapore run into legacy systems and data quality gaps

Take a pull request an agent writes in two minutes that a senior engineer then spends 45 minutes fixing. For software teams, Rao says, the better measure is Mean Time to Release, how long code takes to reach production, weighed against the company’s overall security posture. Code that introduces vulnerabilities or regressions can delay a release rather than speed it up.

In regulated industries, checks once run after deployment are becoming mandatory release gates, and two-minute code that cannot clear them is worth little.

The blind spot: what agents eat

Ask Rao where APAC enterprises are most exposed, and he points not at access but at consumption. Access is increasingly treated as an identity and permissions problem, which security teams broadly understand. Less visible is what agents pull in to finish a job: open-source packages, MCP servers (connectors built on the Model Context Protocol that plug agents into tools and data), skills and plugins. Any of these can carry vulnerabilities or malicious code into enterprise infrastructure before anyone notices.

“The critical control point is where an agent consumes a component,” he says.

When that control fails, the blame does not get diluted. If an agent pulls a compromised MCP server into production at a fintech regulated by the Monetary Authority of Singapore, Rao says the institution deploying it remains accountable, not the model provider or the governance vendor. That means being able to reconstruct which agent acted, under whose authority, what it consumed and why it was allowed. Many firms cannot do that yet: a Sumsub study found Singapore companies’ AI audit trails remain thin.

The vendor in the room

There is an obvious catch. JFrog sells exactly the controls Rao describes, through its MCP Registry, AI Catalog and Agent Guard. “Agents are risky, so buy our governance” is a convenient story for a vendor to tell.

Rao’s defence is that headcount no longer tracks risk. A 30-person startup running agents may operate workflows once associated with a far larger engineering team. If it builds for a bank or insurer, it inherits that client’s security and regulatory demands regardless of size. He concedes startups do not need bank-grade architecture, but argues the minimum, verifying what agents consume, limiting permissions and keeping evidence of how software was built and approved, becomes a commercial requirement long before any breach.

Pricing is harder to pin down. JFrog does not publish list prices for AI Catalog or Agent Guard; Rao says they are integrated capabilities within the JFrog Platform, not standalone products. That may be so, but it leaves buyers calculating governance ROI without a visible line item, an awkward gap for a company urging CFOs to measure cost per outcome.

JFrog is not alone in this race. Sonatype and Snyk have long pitched software supply chain security, and Snyk bought agent-security startup Invariant Labs in 2025. GitHub and Docker run their own MCP catalogues inside platforms developers already live in. For Southeast Asian teams, the question is less whether to govern agents than whose platform gets to hold the keys.

Singapore’s next exam

Rao expects Singapore to get agent governance right first, citing its regulatory track record and concentration of multinationals in regulated sectors. JFrog’s own 2026 Software Supply Chain State of the Union research found 71 per cent of Singapore respondents carefully review AI-suggested security fixes, the highest of eight markets surveyed.

Also Read: SEA’s insurers face a new question: what happens when customers have agents?

The same data shows the cracks. Some 18 per cent have policies against unauthorised AI tools but no way to detect violations, and 59 per cent wait a week or more for open-source package approvals. That is fertile ground for Shadow AI, employees quietly bringing in tools without IT’s knowledge, which Rao says worries organisations more than any rush to hand agents critical decisions. The one call he would keep with humans: which sensitive systems an agent may touch, and how far its permissions stretch.

The lesson from the IVR years still holds. Automation that merely made old processes cheaper disappointed; automation that rethought them paid off. Singapore’s next test, Rao says, is turning governance rules into automated enforcement inside developer workflows. Until then, the agent that writes code in two minutes may still be the slowest route to production.

The post Tokens are the wrong yardstick for AI agents, says JFrog’s Sunny Rao appeared first on e27.