Defending Treasury Against Deepfake Attacks: Practical Low‑Tech Controls

Pretend executives on a video call ordered HK$200 million in wires. The victims followed the instructions.

According to PurpleSec’s investigation into the Arup incident, attackers used synthetic audio and video of multiple executives to push an urgent treasury decision in January 2024. The campaign led to roughly 15 wire transfers totaling about HK$200 million (≈ $25.6 million). PurpleSec reported that, as of early 2025, the funds had not been recovered. That single case shows how convincingly modern generative tools can be used to bypass trust built on sight and sound.

Why look-and-listen verification no longer cuts it

Two realities make this a board‑level problem. First, attackers are already weaponizing synthetic media in real fraud campaigns. PurpleSec’s reconstruction of the Arup flow, which combined open‑source reconnaissance, voice and face synthesis, and a staged meeting that created urgency, reads like a blueprint for social‑engineering at scale.

Second, people aren’t reliable deepfake detectors. A University College London study (reported August 2023, PLOS ONE) found participants correctly identified artificially generated speech only about 73% of the time. Training produced only marginal gains. Other literature reviews and benchmarks show wide variability in human performance depending on media quality and experimental setup. Audio and video alone are no longer definitive evidence of identity.

Automated detectors help, but they aren’t a cure-all. Detection models trained to spot statistical artifacts or compression inconsistencies often perform worse outside controlled lab conditions, and synthesis techniques keep improving. U.S. cybersecurity agencies and industry researchers have flagged limits to artifact‑based detection approaches and recommend treating such tools as one signal among many, not a single safety net.

“Seeing and hearing someone is no longer proof they are real, “ says Deepak Gupta, technical CEO at GrackerAI.

Start with deliberately low‑tech, high‑control gates

The most resilient defenses today aren’t more neural nets chasing other neural nets. They’re controlled, hard‑to‑automate checkpoints that raise the operational cost for attackers and restore verifiable authority to high‑risk actions.

For treasury, HR, and other sensitive workflows, adopt a layered “low‑tech first” baseline:

Phishing‑resistant hardware MFA (FIDO2 / PIV): protect accounts that can approve transactions with FIDO2 tokens or government PIV cards. These are based on public‑key cryptography and block credential‑phishing at the protocol level.
Role‑scoped spoken passphrases: require a randomly generated spoken codeword or phrase tied to the role (e.g., “treasury approver” pool) before executing high‑risk actions in a live call.
Out‑of‑band verification: validate requests over an independent authenticated channel, such as an enterprise app push or an organization‑verified phone number, rather than the same call or email that carried the request.
Dual authorization: require a second authorized approver for first‑time payees, high‑value transfers, or changes to payment instructions.
Secure vaulting and evidence‑based rotation: store passphrases and secrets in an enterprise vault with strong audit logs, require hardware MFA to retrieve them, and rotate only when there’s evidence of compromise, role change, or offboarding (consistent with NIST guidance favoring event‑driven rotation over arbitrary cadence).

These controls are designed to be low friction for legitimate users but high friction for remote attackers. A deepfake can mimic voice and face, but it cannot present a hardware key, complete an authenticated out‑of‑band callback to a verified enterprise device, and recite a role‑specific secret at the same time, unless the attacker already has a foothold inside the organization.

“The defense against the most advanced AI attack is often deliberately low‑tech, “ says Deepak Gupta. James Scobey, CTO at S2i2, underlines the operational point: “Make the control automatic and no‑exception.”

How to make spoken passphrases useful (and not a burden)

Passphrases work when they’re unpredictable, scoped, and managed so staff don’t bypass them under pressure. Practical parameters to follow:

Use randomness from vetted lists: generate passphrases from a vetted wordlist (or a secure generator) rather than letting people invent memorable phrases. A common operational approach is 3-4 randomly selected words (think Diceware‑style selection) which balances memorability and unpredictability.
Scope by role and transaction: assign different passphrases to treasury approvers, HR interviewers, and contract signers. Avoid one‑phrase‑for‑all; a scoped secret limits fallout from a single compromise.
Protect distribution: store passphrases in an access‑controlled vault; require hardware MFA to retrieve them. Never transmit passphrases over email or chat in the clear.
Rotate on cause, not schedule: follow NIST SP 800‑63B’s guidance to replace secrets upon evidence of compromise, role change, or offboarding rather than on a fixed 90‑day cadence.
Audit every use: log retrievals, who requested the passphrase, and the out‑of‑band verification used. If a passphrase is spoken on a call, require the caller to complete a second channel verification before any action.

James Scobey captures the point: “The strength comes from the randomness of the selection.”

Operational edge cases and how to handle them

Every control introduces operational failure modes. Anticipate them and build rules before they’re stress‑tested during an incident.

SIM swaps and SMS callbacks: SMS is weak for out‑of‑band verification. Prefer app push notifications, enterprise telephony tied to directory records, or verified desk lines. If using phone callbacks, require the callback number to be the one on file in HR or corporate directory and log the verification.
Lost or stolen hardware keys: have an attested provisioning and recovery process. Suspend the missing key, require in‑person or biometric re‑provisioning, and establish an emergency recovery path that itself requires multi‑person approval.
Insider coercion or exfiltration of passphrases: limit scope (role‑based, not person‑based), monitor uncommon retrievals, and require a second approver for high‑impact actions. Treat anomalous access to passphrase vaults as a high‑priority incident.
Provisioning and lifecycle: manage hardware keys and passphrases through centralized inventory and attestation. Revoke credentials during offboarding promptly and require re‑attestation when an employee changes role.

A practical pilot plan for treasury teams

Don’t flip company‑wide controls overnight. Start small, measure, and expand based on metrics.

Scope: pilot with the top 3 treasury users and the 20 highest‑value payees for 60 days.
Controls to enable: require FIDO2 hardware keys for approvers, role‑scoped spoken passphrases stored in a vaulted manager, out‑of‑band enterprise callback, and dual authorization for flagged transactions.
Example thresholds: require dual authorization for first‑time payees and for transfers above $50, 000 (adjust to your risk profile and regulatory obligations).
KPIs to track: average time‑to‑approve, percentage of transactions requiring exceptions, number of blocked or reversed fraudulent attempts, user satisfaction (qualitative), and audit‑log completeness.
Success criteria: minimal operational friction (approval time within acceptable range), reduction in exceptions, and proof that the controls would have stopped past incidents discovered during analysis.

Where these defenses fall short (and what to plan for)

No single layer is foolproof. Expect attackers to combine synthetic media with account compromise, social pressure, or insider collusion. Plan for these realities with specific mitigations:

Combined attacks: deepfakes plus stolen credentials are more dangerous than deepfakes alone. Require multi‑factor step‑ups for actions that change payment rails or create new privileged accounts.
Supply‑chain and provisioning attacks: protect the hardware key lifecycle and the vaulting system with strong vendor controls, attestation, and regular audits, and don’t assume vendor defaults are sufficient.
Human factors: make controls automatic and blanket. When staff experience controls as consistent rather than targeted, they’re less likely to create exceptions during stressful calls.

Key takeaways: quick questions with honest answers

Are live audio and video still reliable proof of identity?

No. Multiple investigations and a University College London study show humans miss a meaningful share of synthetic speech and that modern models can produce convincing impersonations. Treat AV alone as insufficient for high‑risk actions.
Will automated detectors save us?

Not on their own. Artifact‑based detection can help as one signal, but agency advisories and industry research warn that such tools should not be the single line of defense. Use detection as part of a layered strategy.
Which control should I prioritize first?

Phishing‑resistant authentication for any account able to approve transactions, FIDO2 or PIV tokens, is the highest‑leverage starting point. Pair that with automatic step‑ups (no‑exception) for high‑risk workflows.
How should passphrases be managed at scale?

Generate them randomly from vetted lists, scope them by role/transaction, vault them in an access‑controlled manager, require hardware MFA to retrieve, and rotate on evidence (compromise or role change), not on a fixed calendar.
Will these measures be usable for staff?

They add intentional friction. That’s the point. Keep controls consistent and automated; pilot small, measure approval times and exception rates, and tune thresholds to balance security and operations.

Take the pragmatic path, not the physics race

Chasing each new generative model with another detector is a losing arms race. The immediate, practical win for organizations is to restore authority to controls attackers can’t reproduce from a short audio clip: hardware keys, authenticated out‑of‑band checks, scoped spoken secrets, and enforced dual approval. Those deliberately low‑tech checkpoints raise the cost of success for attackers dramatically and buy time, time to detect, disrupt, and adapt as synthetic media evolves.

Sources cited in this piece include PurpleSec’s investigation of the Arup incident and a University College London (PLOS ONE) study on human detection of synthetic speech; NIST SP 800‑63B provides recommended practices on credential rotation and evidence‑based change.