Singapore signs an AI safety MOU with Anthropic and builds an on-device scam filter with Google DeepMind

What does Anthropic's Singapore MOU actually commit anyone to? Reading the October 14, 2026 announcement alongside Google DeepMind's same-day post about a Singapore scam detector, the answer is: a cooperation framework from Anthropic and a working prototype from Google, with money and dates absent from both.

Both were unveiled around Singapore International Cyber Week, and both show a small country collecting commitments from competing labs at once.

The Anthropic MOU

Anthropic opened its fifth office in the Asia-Pacific region and signed a memorandum of understanding with the Ministry of Digital Development and Information (MDDI). Kai Fong Chng, MDDI's Permanent Secretary, and Anthropic co-founder Ben Mann signed it. Minister Josephine Teo and Michael Sellitto, Anthropic's Head of APAC Policy, witnessed it.

It covers three areas: AI safety, security and model evaluation; research on AI adoption and its effects; and responsible adoption across the public and private sectors. The listed actions are technical exchanges with Singapore's AI safety and assurance institutions on evaluation, testing standards and incident response; AI skills for public officers; help for small businesses adopting AI; sharing Anthropic Economic Index research; and policy expertise as Singapore develops its AI rules.

An MOU is a statement of intent. It carries no binding obligation and, as e27 points out, no investment figure, headcount target or funding for the SME and skills programs.

The numbers behind the pitch

Anthropic supports the deal with usage data that is mostly its own. In its latest Economic Index, Singapore ranks second of 121 countries for Claude.ai usage relative to population, at nearly six times what its size alone would predict. More than 1,500 GovTech Singapore engineers and technical staff use Claude Code. Grab says its Claude-based sales assistant saves about 40,000 hours a quarter.

e27 adds a useful caution: usage per head says little about value created, and the customer figures come from Anthropic and its customers, not independent audits.

The Google DeepMind scam model

The same day, Google DeepMind and GovTech Singapore described a privacy-preserving model that detects scams on a phone. A lightweight Gemma model was fine-tuned on local scam conversations, including impersonations of banks, public institutions and local figures. It runs in two stages: a lightweight classifier screens incoming messages with minimal power, and a fine-tuned model analyzes a larger conversation snapshot to confirm or dismiss a threat.

The problem is large. DeepMind says investment fraud, fake job offers and impersonation of officials make up over half of Singapore's reported crimes, and that scams have cost more than S$4 billion since 2019. Social engineering accounts for up to 80% of losses.

On privacy, the model runs entirely on the device, so conversations are not scanned on servers, and neither the Singapore government nor Google can access them through the deployment. Fine-tuning took place in GovTech's managed Google Cloud environment with anonymized data. Mark Chen of GovTech put the design goal this way: "Fighting scams shouldn't mean sacrificing personal privacy."

Has it shipped?

This is where the reporting splits. DeepMind says the rollout will be phased and developed with the Singapore government. One outlet described the model as shipped, while another said no public release date, supported messaging apps or independent accuracy figures had been announced, and that it is meant for messaging platforms and operating system makers to integrate. The DeepMind post itself gives no accuracy or false-positive numbers.

For a scam filter, the false-positive rate is the whole product. A tool that warns users too often gets switched off. Until someone publishes it, the claim is an architecture, not a result.

The competitive picture

e27 frames the Anthropic deal as both a safety partnership and a sales operation. It notes OpenAI hired a new Asia-Pacific sales head, Google is scaling Gemini through schools and telcos, and Singapore has its own local SEA-LION model as an alternative. It also flags a dependency risk: in June, Anthropic suspended access to its Fable and Mythos models for nearly three weeks to comply with US export controls, and the MOU does not address that.

That risk is why the sovereignty debate matters here. We have written about what the phrase means in practice in our explainer on sovereign AI.

What Singapore gets, and what it does not

Singapore's interest is clear. It wants frontier labs to test their models against local conditions, to help its public service build skills and to give its institutions a seat in how evaluation standards develop. Anthropic took part in the Generative AI Evaluation Sandbox run by IMDA and the AI Verify Foundation in 2023, and later ran a red-teaming project with them in English, Mandarin, Malay and Tamil. The MOU extends a relationship that already existed.

What Singapore does not get is a commitment on supply. The MOU as described lists cooperation, not guaranteed access, and Anthropic's announcement does not mention service continuity. For a government that already runs engineers on Claude Code, that is the gap to close.

Our take

Singapore is doing the sensible thing for a state of its size: signing with several labs and keeping its own models alive. The two announcements are not equal. The DeepMind project solves a measured national problem and ships a concrete design. The Anthropic MOU is the more interesting for safety policy, because it promises exchanges on testing standards and incident response, and those could become useful if Singapore publishes them.

What we would watch: published evaluation standards from MDDI or its assurance institutions, a named release for the scam model with accuracy data, and whether the MOU ever says anything about continuity of access.