What is the HKMA GenA.I. Sandbox?

A Hong Kong Monetary Authority program, run with Cyberport since 2024, giving financial institutions a controlled environment to test AI use cases; its second cohort covers 27 projects from 20 banks and 14 technology partners, including deepfake detection trials.

What deepfake detection results did Hong Kong banks report?

Ping An Digital Bank, with OneConnect, reports above 96 percent detection of manipulated and synthetic selfies in identity verification; HSBC reports that roughly 20 tested deepfake attempts all failed to bypass its controls, with testing methods not publicly detailed.

How should a 96 percent deepfake detection rate be interpreted?

Alongside its missing parameters: which attack types and generators, the false-positive rate at production volume, and whether tests included injected streams and novel generators. Detection scores without methodology are promising starts, not procurement facts.

Why are regulators hosting deepfake testing sandboxes?

Because no single institution can maintain a current attack corpus: shared infrastructure, pooled attack classes, and comparable trials make defense measurable, a pattern spreading as regimes like Europe's push attack testing toward certification.

Does deepfake detection at onboarding stop executive impersonation fraud?

Only partly: detection guards monitored channels like video KYC, while the costliest impersonation chains arrive over personal messaging and cloned voice calls, which is why payment ceremonies that no convincing media can waive ride beside detection in durable programs.

Back to News

The Deep Brief · Oct 19, 2026 · 4 min read

Hong Kong banks post first deepfake-detection sandbox scores

HKMA's GenA.I. Sandbox posts early deepfake defense results: Ping An tops 96% detection and HSBC repelled all tested attempts, with methods unpublished.

Rosalie Chirip

Senior Editor at deepidv

Hong Kong's Central skyline across Victoria Harbour, with the Two IFC tower that houses the HKMA rising above the banks.

The first regulator-hosted deepfake defense scores are in, and they arrive from the city that gave the threat its founding case. Hong Kong banks are reporting results from AI deepfake detection trials run inside the Hong Kong Monetary Authority's GenA.I. Sandbox, the Cyberport-hosted program whose second cohort spans 27 projects from 20 banks and 14 technology partners, selected from more than 60 proposals, with trials running since early 2026. The motivating memory is local: the 2024 Arup case, $25 million wired after a video call of deepfaked executives, happened in Hong Kong, and HKMA deputy chief executive Arthur Yuen frames the cohort as "a major leap forward" in making AI adoption safer.

Two results headline. Ping An Digital Bank, partnered with OneConnect Financial Technology, reports its Anti-Fraud Strategy Platform flags manipulated and synthetic selfies during identity verification at better than 96 percent detection accuracy. HSBC tested an unnamed detection system against roughly 20 deepfake attempts and reports none bypassed its controls, alongside disclosing a 150-strong Hong Kong AI team it plans to grow. The cohort's breadth beyond deepfakes is itself notable, relationship-manager copilots cutting hours to minutes at Standard Chartered, sign-language translation at HSBC, with governance features, automated output checks, fraud-resilience simulations, built into the sandbox's design.

Reading the scores like the testing era demands

The results deserve both welcome and the scrutiny the article itself invites: limited public detail on methods makes the systems hard to compare or extrapolate. The questions that convert a sandbox score into a procurement fact are the standard ones from detection accuracy's fine print: detection of what, presented artifacts, synthetic selfies, injected streams, against which generator families, at what false-positive rate, on what base rate of attacks? A 96 percent catch rate is excellent or inadequate depending entirely on the false-alarm cost at millions of genuine sessions, and twenty repelled attempts is a strong anecdote whose power depends on who built the twenty and how adversarially. None of this diminishes the trials, sandboxes exist to generate exactly these first numbers, but the hardening testing era has set the bar for what counts as evidence: published attack taxonomies, false-positive rates, and holdout generators, the methodology that separates measurement from demonstration.

Why the sandbox pattern will travel

The structural news is the venue: a central bank hosting adversarial AI testing, with shared infrastructure at a supercomputing center, a Collaboratory pairing banks and vendors, and resilience simulations as standing features. That pattern answers the problem every institution faces alone, no bank can maintain a current attack corpus by itself, and it will travel: regulators from Singapore to London run adjacent sandboxes, and the EU's testing deadlines are pushing attack testing from optional to certified. For banks elsewhere, the Hong Kong cohort is a preview of the ask coming their way: demonstrable deepfake defense at the video-KYC and onboarding gates, measured against live attack classes, with the quiet lesson of the quarter attached, detection guards the monitored channels, while the Fideuram chain arrived on an executive's personal phone, which is why the ceremony layer rides beside the classifier in every durable architecture. Scores open the conversation; methodology and ceremonies finish it.

HKMA Deepfake Sandbox FAQ

What is the HKMA GenA.I. Sandbox?: A Hong Kong Monetary Authority program, run with Cyberport since 2024, giving financial institutions a controlled environment to test AI use cases; its second cohort covers 27 projects from 20 banks and 14 technology partners, including deepfake detection trials.
What deepfake detection results did Hong Kong banks report?: Ping An Digital Bank, with OneConnect, reports above 96 percent detection of manipulated and synthetic selfies in identity verification; HSBC reports that roughly 20 tested deepfake attempts all failed to bypass its controls, with testing methods not publicly detailed.
How should a 96 percent deepfake detection rate be interpreted?: Alongside its missing parameters: which attack types and generators, the false-positive rate at production volume, and whether tests included injected streams and novel generators. Detection scores without methodology are promising starts, not procurement facts.
Why are regulators hosting deepfake testing sandboxes?: Because no single institution can maintain a current attack corpus: shared infrastructure, pooled attack classes, and comparable trials make defense measurable, a pattern spreading as regimes like Europe's push attack testing toward certification.
Does deepfake detection at onboarding stop executive impersonation fraud?: Only partly: detection guards monitored channels like video KYC, while the costliest impersonation chains arrive over personal messaging and cloned voice calls, which is why payment ceremonies that no convincing media can waive ride beside detection in durable programs.