The Singapore AI Safety Institute is the designated national AI safety institute for Singapore. Its stated purpose is to "drive Singapore's AI evaluation and testing work, to support addressing gaps in global AI safety sciences" (Source: sgaisi.sg). It is not a standalone body: the technical work is carried out by the Digital Trust Centre (DTC), a national research centre hosted at Nanyang Technological University, in partnership with the Infocomm Media Development Authority (IMDA), which acts as the overall government agency for the institute and fronts its international engagements. Singapore designated the DTC as its AI Safety Institute in 2024 (Source: imda.gov.sg).
The DTC was set up in 2022 and is funded by a S$50 million grant from the National Research Foundation, Singapore and IMDA, running from October 1, 2022. Its four stated focus areas are trust-tech research, trust-tech innovations, a sandbox environment for businesses experimenting with trust technologies for data-sharing challenges, and nurturing 100 research-and-development talents in digital trust (Source: sgaisi.sg).
Technical testing
The institute names two priority areas for its own testing.
Multilingual and multicultural testing aims to surface failures that appear only in non-English settings and to check that safety measures hold for minority groups. Reported work includes multilingual agentic testing using Project Moonshot, the open-source evaluation tool of the AI Verify Foundation, carried out as part of joint testing under the Network for Advanced AI Measurement, Evaluation and Science (NAAIMES); multilingual translation of local safety benchmarks with the Japan AI Safety Institute; and an evaluation of large-language-model judge consistency across languages with the Canada AI Safety Institute (Source: sgaisi.sg).
Agentic AI testing studies how agentic behaviours emerge and scale, including an evaluation of data-leakage risks in routine agentic tasks conducted with the Korea AI Safety Institute (Source: sgaisi.sg). See Agentic AI.
A third strand, societal impact studies, examines effects on individuals and institutions including emotional dependency and cognitive offloading.
Double-blind evaluation pilot
In August 2026 the Singapore AISI was one of two evaluator partners in the double-blind evaluation pilot published by Google DeepMind with AVERI, MLCommons and OpenMined. Gemini 2.5 Flash Lite was evaluated against a private prompt set focused on harmful-content elicitation in Singapore's context, inside a hardware-encrypted enclave in which neither the evaluator's prompts nor the model's weights were visible to the other party. Akriti Vij is listed as the institute's author on the technical report (Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing). The Singapore-context prompt set is an instance of the multilingual and multicultural testing priority described above, run under a confidentiality arrangement that keeps the prompts reusable. See Double-Blind Evaluation.
International position
The institute participates in the international network of AI safety institutes, listed there under Singapore, and describes joint testing exercises with NAAIMES members alongside regional capacity-building in ASEAN and the Asia-Pacific through workshops and red-teaming challenges, and contributions to the International Scientific Exchange on AI Safety (Source: sgaisi.sg).
Relationships
- related: Infocomm Media Development Authority (IMDA, Singapore) — the overall government agency for the institute.
- related: International AI Safety Institute Network (INSAI / AISIN) — member of the international network of AI safety institutes.
- related: Double-Blind Evaluation — evaluator partner in the 2026 pilot.
- related: Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing — technical report co-authored by the institute.
- related: AI Benchmarks and Evaluation — multilingual and agentic testing methodology.
- related: AI Verification and Evaluation Research Institute (AVERI), OpenMined — pilot partners.