As AI-driven fraud grows and regulations diverge, executives must establish consistent security standards while adapting controls to local legal requirements.
In January 2024, a finance employee at Arup’s Hong Kong office joined a video call with the group’s apparent CFO and colleagues. Every participant was a deepfake. Police said he authorized 15 transfers worth about US$25.6 million. Arup said no internal systems were compromised. The failure was in verification, not infrastructure. That vulnerability is reflected in breach data across regions, even as regulators from Brussels to Seoul adopt diverging rules. For multinational leaders, the question is how to build one defensible standard across jurisdictions that disagree.
From perimeter defense to identity assurance
Enterprise security was designed for threats that entered through systems. AI shifted the target to identity, instruction and data. IBM’s 2026 Cost of a Data Breach Report put the global average at $4.99 million, up 12%. One in four malicious breaches was AI-enabled, costing about $6 million each. Deepfake impersonation made up 45% of AI-enabled attacks, the Arup pattern. Among organizations with AI-related breaches, 92% lacked proper AI access controls. The techniques are new; the control failures are not.
Boards are moving ahead of incidents. Some 85% of surveyed organizations plan higher security spending after learning of frontier AI capabilities, against 64% after an actual breach. The shift suggests greater emphasis on anticipating threats rather than responding after losses occur.
Costs and rules diverge by region
The United States carries the highest cost in IBM’s country ranking: $11.5 million per breach, up 13%. At more than double the global average, it sets the ceiling for loss modeling. Financial services and energy bear the brunt of AI-related attacks, at $6.29 million and $5.2 million per breach.
Latin America averaged $4.65 million, with about 19% of malicious attacks AI-enabled. ASEAN averaged $4.12 million. Lower reported AI involvement in Latin America should not be read as evidence of lower risk; reporting and adoption patterns may differ.
India set a record at ₹25.5 crore, up 15.9%, with shadow AI, regulatory non-compliance and cloud migration the top cost drivers. Organizations without AI and automation averaged ₹31.6 crore. The Middle East averaged $8 million, about 60% above the global figure. Financial services and technology each reached $10.67 million, with excessive privileges, mismanaged secrets and weak role management among the leading contributors. Across both markets, weaknesses in governance and access management compound financial exposure.
In Africa, Interpol found AI involved in 55% of reported cybercrime cases. Losses rose from $192 million to $484 million in a year. South Africa recorded the sharpest percentage increase in IBM’s survey, 22%, to $3.04 million. Reporting is uneven, so the 55% figure should be treated as indicative, not comprehensive. The figures point to a widening challenge for enforcement as fraud scales.
Rules diverge as sharply as costs. The EU’s AI Omnibus, in force since 27 July, moved high-risk obligations to December 2027 and August 2028. A US executive order of 11 December 2025 targets state AI laws through litigation, yet those laws stay enforceable until courts invalidate them. China has mandated explicit and embedded labels on AI-generated content since 1 September 2025. South Korea’s AI Basic Act took effect on 22 January 2026 with extraterritorial reach, while Japan’s AI Promotion Act carries no penalties. These five regimes reflect distinct enforcement approaches, leaving multinational companies without a common regulatory framework.
Each risk has a matching response
Fragmented regulation pairs with a single control baseline. The Omnibus leaves the Act’s risk-based architecture in place, so it alters timing, not scope. US state laws remain in force during litigation, and Korea’s law reaches foreign providers. Building to a strong common standard and adding local legal requirements can reduce the need to redesign controls for each market. The deferral creates engineering time, not a reason to delay preparation.
Identity weakness pairs with identity-centered defense. Extensive security AI and automation cut breach costs by $1.93 million globally and by more than $3 million in the Middle East, yet 23% of organizations there had not adopted it. Those savings exceed the roughly $1 million additional cost associated with AI-enabled attacks, before accounting for tool costs. The average breach lifecycle rose to 247 days, its first increase in five years. Automation can help address longer detection and response cycles, while privileged-access controls target a cost driver identified in regional data.
Capacity gaps pair with public-private cooperation. The World Economic Forum’s Global Cybersecurity Outlook 2026 flags a severe talent shortage and fragmented defenses. No enterprise can out-hire an attacker that automates. Seventeen African countries updated cybercrime laws in 2025, indicating progress in legal capacity. Coordinating among banks, telecom companies, and police offers a practical starting point.
Set one standard, localize the rest
Boards should read regional breach data as one exposure, not several. CFOs and CISOs should fund identity assurance and payment verification ahead of regional compliance builds. General counsel should treat divergent regimes as overlays on a common baseline. Policymakers should note where losses grow fastest: where reporting and coordination are weakest.
Regulators will not converge soon; attackers already operate without borders. Leaders can prepare by establishing a common security standard, applying it across the organization and adapting it where local law requires.
