Your next audit won't ask about deepfakes. Your next attacker will.

In March 2025 a finance director in Singapore was contacted on WhatsApp by someone posing as the company's chief financial officer, then joined a Zoom call with the chief executive and other senior colleagues. According to Singapore Police, the executives on that call were deepfakes. The director transferred more than US$499,000 before police in Singapore and Hong Kong stepped in.¹

The attack never touched the inbox. It started in a messaging app and ended on a video call. Most security awareness programmes, and most of the rules that shape them, still start and end with email.

If most firms are still training for the inbox, are the regulators any further ahead? We wanted to know. So we read the rules and guidance that set expectations for staff security awareness in North America, Asia-Pacific and the Gulf, with the EU as a reference point, and scored each document on nine questions: does it require training, require testing staff, name phishing, name voice or SMS attacks, name AI or deepfakes, require measuring whether training works, set a frequency, require board reporting, and require staff to verify payment requests through a separate channel. Every non-zero score is backed by an exact quote. An independent reviewer then re-checked every binding rule against its source.

Training is common. 20 of the 29 binding rules in force mention staff training. New York's Part 500 requires annual training "that includes social engineering for all personnel".² The US banking agencies' standard, written in 2001, asks firms to "train staff to implement" their information security programme.³ The EU's DORA makes security awareness a compulsory part of staff training, yet the words "phishing" and "social engineering" do not appear anywhere in it.⁶

Testing people is rare. Only 3 binding rules specifically expect firms to test how staff respond, and only one, Kuwait's Cyber and Operational Resilience Framework of December 2025, makes phishing simulation mandatory.⁷ Australia's CPS 234 and Singapore's cyber hygiene notice, two of the most cited rules in the region, say nothing about staff training at all; in both markets, the people controls sit in non-binding guidance.^4,5

AI barely appears. Two binding rules name AI-enabled attacks: Kuwait's framework and Malaysia's Risk Management in Technology policy, updated in September 2026. Both mention AI in their fraud-detection clauses, not in what staff are trained or tested on.^7,8 And not one of the 29 rules requires staff to confirm a payment instruction through a separate channel, the single control that stops a deepfake "CFO" on a video call.

The threat is not missing from regulators' thinking. It is in a different kind of document. Of the 28 documents in our set that name AI-enabled or deepfake attacks, 26 are guidance, alerts, letters or information papers. Every one of them was published in or after June 2024.

Some of that guidance is precise. New York's Department of Financial Services told firms in October 2024 that training on deepfake attacks "can be effectively delivered through simulated phishing, and voice and video impersonation exercises", and that staff should "verify a requestor's identity and the legitimacy of the request if an employee receives an unexpected money transfer request by telephone, video, or email".⁹

The Monetary Authority of Singapore's 2025 paper goes furthest, describing deepfake simulations, call-back "through a separate and trusted communication channel", code words and dual control.¹⁰ Abu Dhabi's FSRA told firms in 2025 to verify requests "through trusted methods (e.g., phone calls for financial transactions)".¹¹ FinCEN warned US institutions about deepfake media in November 2024.¹² None of these documents is a rule.

That is how regulation usually moves: guidance first, rules later, and rules take years to write. But it leaves a gap today. A bank could meet every binding requirement in our set and never have tested whether its finance team would stop a deepfake call. Even the most advanced rule we found, Kuwait's December 2025 framework, names AI only in its fraud clauses. On this threat, nobody's rulebook is fully ready yet, including the regulators' own. For firms that move first, that is an opportunity, not a worry.

We collected corporate social-engineering incidents disclosed between October 2024 and October 2026. Of 11 cases, 7 began with a phone call, a video call or a WhatsApp message. In January 2026 Swiss police reported that a company owner made transfers worth several million francs after calls from an AI-manipulated voice imitating a known business partner.¹⁵ In 2025 Google and Cisco each disclosed that staff were talked into granting access by attackers on the phone.^16,17 None of the 11 was disclosed as an AI-written email.

The aggregate figures point the same way, with caveats. The FBI's 2025 Internet Crime Report logged 22,364 complaints that mention AI, with losses above US$893 million; that flag is self-reported, so it is a floor, not a measure.¹³ In Singapore in 2025, business email compromise losses fell 60.1% while losses to phone and messaging impersonation of officials, mostly aimed at individuals, rose 60.5%.¹⁴ Europol's 2025 assessment describes the same shift towards voice and mobile channels.²³

Spotting the fake is not a reliable defence. Pooled across 56 studies and more than 86,000 participants, people identified deepfake video about 57% of the time and deepfake audio about 62%, not reliably better than chance.¹⁸ That is why the guidance keeps returning to process: verify through another channel, whatever your eyes and ears tell you.

Programmes built to satisfy the rules mostly test the channel the rules name. We reviewed the product pages of 10 widely used security awareness vendors in October 2026. All 10 simulate phishing email. Six list voice-call tests and five SMS. Three list cloned-voice tests, two list Teams, Slack or chat lures, and two list a deepfake video-call simulation.

Email testing still matters, but on its own it moves the needle less than many assume. In a randomised study of more than 19,500 employees at UC San Diego Health, annual training showed no significant link with failing a phishing test, and training shown after a click reduced failures by only 1.7 percentage points.¹⁹

Firms do not have to guess what good looks like. Regulators in New York, Singapore and Abu Dhabi have already described it in plain language: practise against voice and video impersonation, verify unusual payment requests through a separate channel, and measure whether people actually do it. Several Gulf and Asian rules already require evidence that training works.

A call-back on a known number, a code word and a second approver for urgent transfers are simple controls that need no new technology. Practising them under realistic pressure is what turns them from policy into habit. Firms that do this now will meet the next rule as a formality rather than a project, and, more importantly, will have a team that stops the call when it comes.

1. Test the channels attackers use. Add voice, video and messaging scenarios to email, as NYDFS and MAS describe.^9,10
2. Make verification a payment control, not a training tip. Call back on a known number, use code words, require a second approver for urgent or unusual transfers. MAS's paper and Hong Kong's SFC (maker-checker for high-impact actions) both point here.^10,22
3. Measure behaviour, not completion. Malaysia, Bahrain and Kuwait already require evidence that training works; reporting rates and verification under pressure are better signals than course completions.^7,8,20
4. Give the board people-risk numbers. Bahrain requires an update on staff awareness at every board meeting. That is a sensible default anywhere.²⁰
5. Don't wait for the rule. In most markets in our set, the clearest description of today's attack, and of what good looks like, is already written down in guidance. Meeting the binding minimum is not the same as being ready for the call.