Founded and led by the Hong Kong Association for the Advancement of Science and Technology (HKAAST) · Secretariat provided by HKAAST 繁 · 简 · EN Want to know how the National AI Safety Governance Framework 3.0 lands in your industry?
GBA AI Safety & Development Standing CommitteeNon-profit AI Safety Think Tank
National AI Governance Framework 3.0 · Research × Open Source × Governance × Education

Building the line of defence through research
Making AI Safe, Reliable & Trustworthy

The Greater Bay Area AI Safety & Development Standing Committee is a non-profit think tank founded and led by the Hong Kong Association for the Advancement of Science and Technology. Our work is driven by our own research — from the HoloTSH tensor–hypergraph school and the HoloBlocker topological circuit breaker to split-inference privacy auditing — turning AI governance into computable, verifiable, auditable mathematics and engineering, then giving it back to society through open-source tools, industry–academia partnerships and public education.

🤖 Ask the AI Safety Assistant — free: tools / audit / partnership → Enterprise partnership · Governance adoption → What is Framework 3.0?
10Academic preprint papers
6+University & institute partners
19Corporate member organisations
70+Committee column articles

National AI Governance Framework 3.0: From "Compliance Documents" to "Verifiable Governance"

Framework 3.0 self-check matrix: eight lifecycle stages × four risk categories

In response to the national 15th Five-Year Plan's strategic deployment on AI safety governance, the Committee proposes the "Framework 3.0" implementation pathway — governance cannot stop at policy documents; it must close the loop into a computable, auditable and enforceable technical layer.

📜

1.0 Institutional · Drawing boundaries

Laws, standards and industry codes draw the red lines (Generative AI Measures, EU AI Act, etc.). Necessary — but documents cannot govern each individual inference.

🧭

2.0 Process · Building systems

Organisations establish AI governance committees, risk tiering and model lifecycle management. The system exists — what is missing is enforceable "teeth".

🔬

3.0 Technical · Verifiable Our home ground

Encoding governance requirements as mathematical constraints: auditable outputs (SA-RI measurement), interceptable behaviour (HoloGovernor governance gate), instant circuit-breaking on boundary violations (HoloAegis / HoloBlocker) — the three-line suite is the implementation engine of Framework 3.0.

Research

Our research follows one theme — "mathematically verifiable safety": encoding governance requirements into geometric and topological language so that AI judgments, actions and data flows become computable, auditable and interceptable. Academic status: 10 preprints published, 8 journal papers under review, and the HoloAuditor research dataset published on IEEE DataPort. Six flagship programmes below; details on the Research page and in the publication list.

Architecture · Global launch

AI Safety "Twin Star" Architecture

HoloAuditor (holographic auditor) + HoloBlocker (embodied-AI circuit breaker), launched globally at the 2026 World Young Scientists "AI+" Dialogue in Macao, building a mathematical safety floor for embodied intelligence.

DOI 10.2139/ssrn.6368538 · 6421619 · 6368638
TCM × AI

HoloTSH Tensor–Hypergraph Framework

Translating TCM "pattern differentiation" into tensor manifolds and hypergraph audit, led theoretically by Prof. Ng Kwok-Po (SIAM Fellow) of HKBU, enabling computable, trustworthy and traceable TCM data.

DOI 10.36227/techrxiv.177130202.27796764
Embodied-AI safety

HoloBlocker Topological Circuit Breaker

Reveals the intrinsic risk of "survival motives" in embodied AI and proposes millisecond topological fusing: any behaviour drifting off the safety manifold is physically cut off — better a blown fuse than a disaster.

DOI 10.12074/202604.00098
Content safety

HoloAegis Safety Manifold

Frozen representations + topological inference replace black-box classifiers: a minimal-parameter, zero-shot LLM guardrail that is auditable and self-hostable for open models.

DOI 10.48550/arXiv.2608.08485
Privacy measurement

SA-RI Split-Inference Privacy Audit

A dual-track empirical audit of representation leakage in split learning / split inference — turning "leakage risk" into measurable, comparable, fresh numbers that support data-export decisions.

DOI 10.13140/RG.2.2.32481.67681
Education × Data

ShenNong AI & GBA School Research Network

A "Digital Great Wall" for adolescent health data and a K12 AI-driven research network — the trusted data foundation for Hong Kong's Digital Education Blueprint and the national "Data Element ×" strategy.

DOI 10.12074/202606.00271

Media & Press

Our researchers' views and initiatives are widely covered by media in mainland China and Hong Kong.

People's Daily (Overseas Edition) · 22 Jul 2026 · p.10 Education & Science

"When machines start to think, how should humans get along with them?" (Questions of Our Era ①)

Committee member Dr. Li Tak-Ho (Alex), Executive Dean of the Guangdong Digital Industry Research Institute, on AI's three leaps — capability, goal and speed — noting that "AI iteration is exponential while legislation and ethics discussion is linear".

"Today's human–AI relationship resembles that of a baby with immature parents." — Li Tak-Ho
Read the People's Daily article →  More coverage →
Wen Wei Po · 18 Jan 2026 · Full A4 pageBuilding a TCM Data International Port"Data does not cross the border — compute does": the XDataTrust platform → Ta Kung Pao · 30 Mar 2026Four universities launch the "Twin Star" AI-safety architectureA global first, a mathematical safety line for embodied AI → HK01 In Depth · 22 Sep 2026The rise of China's compute ③: a "Data Customs"Interview on "data does not cross the border, compute does" → People's Daily Overseas · 22 Jul 2026When machines start to think…The "baby and immature parents" interview →

Open-Source Tools (our research outcomes)

All tools below are incubated by the Committee's research programmes, under the principles of openness, auditability and self-hosting, and now form a three-line toolkit: Measurement (SA-RI) · Governance (HoloGovernor) · Screening (HoloAegis). Source code and whitepapers are released on an "open after academic publication" schedule; items not yet public will follow within about two weeks.

Guardrail

HoloAegisResearch preview

Zero-shot LLM content-safety manifold: frozen representations + topological inference deliver an auditable guardrail with alert mode.

Live demo environment (demo account via the Secretariat) · demo pack 3.2MB · deployed P50 26.3ms →
Circuit breaker

HoloBlockerResearch preview

Embodied-AI topological circuit breaker: millisecond physical cut-off the moment behaviour leaves the safety manifold.

Technical docs (GitHub) →
Geometric audit

HoloAuditorResearch preview

Maps domain axioms (e.g. TCM "monarch–minister–assistant–envoy") onto mathematical manifolds, intercepting "popular but wrong" AI output in 0.82 ms.

Technical docs (GitHub) →
Privacy audit

SA-RI Audit EngineSample report available

Split-inference representation-leakage measurement: dual attack families, cross-metric empirical audit, AMBER/RED-graded export-safety reports. Core engine v0.3.x is open source (GitHub).

Open-source engine v0.3.x (GitHub) →  📄 Sample report →
Data space

Trusted Data Space SuitePartner co-build

Onboarding & listing dual review, controlled-task registration and metering/reporting for industrial data spaces, aligned with the "Data Element ×" strategy.

Public education

AI Safety AssistantOpen to all

The Committee's public-education chatbot built on our knowledge base — free, bilingual, answers instantly.

Try it now →

What "3token" means

"3 locks" is our core attitude — every use of AI deserves to be guarded by three criteria: Safety, Reliability, Trustworthiness.

Safe · 🔐

Know the risks of AI — phishing, privacy leaks, deepfakes, data leaving your borders. Learn to recognise and defend, so technology never becomes the entry point of harm.

Reliable · ⚙️

Understand that AI can err and hallucinate at any moment. Verify sources, use trustworthy tools, minimise data — build decisions on solid ground.

Trustworthy · 🤝

Know where your data goes, how it is stored and shared. Assert data sovereignty, transparent governance and ethics — take back control of your data.

Public AI Literacy

Plain-language AI-safety content and a free assistant for the general public — a dedicated column alongside our research line.

Guide · Beginner

Five everyday AI pitfalls and how to protect yourself

Fake customer service, phishing links, over-permissioned apps and more — the most common AI risks and the defences you can apply today.

Read →
Guide · Intermediate

Data export & cross-border AI: the law and the tech

What happens when your data is processed abroad? Data minimisation, localisation and consent — three concepts explained at once.

Read →
Free · Instant

AI Safety Assistant

Questions about phishing, passwords, privacy, deepfakes or data export? Ask our education assistant — free and non-commercial.

Ask the assistant →

About us

The GBA AI Safety & Development Standing Committee was founded and led by the Hong Kong Association for the Advancement of Science and Technology (HKAAST, a non-profit professional body founded in 1985), responding to the national 15th Five-Year Plan's strategy for self-reliant science and AI-safety governance, together with the GBA research alliance, Huawei, Linux Foundation APAC and corporate members.

The Committee connects a cross-disciplinary research network spanning the University of Macau, Hong Kong Baptist University, The Education University of Hong Kong, Guangdong University of Technology and the Guangdong Digital Industry Research Institute — launching the "Twin Star" AI-safety architecture globally at the World Young Scientists events and issuing the GBA Joint Initiative on AI Safety in Education.

HKAAST is an institutional member of the World Young Scientists Alliance (WAYS); the Committee connects to the young-scientist network through it.

As a non-profit think tank, we are research-first, open-source and education-first. Income from assessments and governance services supports public-interest research and education.

Committee structure & members →

What we do

Contact us

For research collaboration, service enquiries, media interviews or open-source co-development, reach the Secretariat any time.

📧 info@3token.org

Or simply ask the AI Safety Assistant — free and open to the public.