Introduction

AI-powered cyberattacks have moved from theory to daily operations. Attackers now use generative models to write phishing emails, clone voices, and produce deepfake video. They automate the search for software flaws and scale fraud across borders. For businesses in 2026, this shift is not only a security problem. It is a legal problem. Courts, regulators, and insurers are starting to ask a hard question. When an AI-enabled attack causes loss, who pays?

This blog explains the new legal risk. It covers the threat landscape, the rules that now apply in the European Union and the United Kingdom, the liability gaps that expose firms, and the steps that reduce exposure. The aim is practical. Directors, compliance teams, and risk officers need a clear view of where duty of care now sits.

The 2026 Threat Landscape

The volume and quality of AI-driven attacks have risen sharply. Deepfake incidents increased by more than 2,000 percent between 2022 and 2025. Voice cloning and synthetic video now support business email compromise and fraudulent payment requests. AI-generated phishing cuts the cost of an attack and raises its success rate, because messages read as genuine and target victims at scale.

Financial services face the sharpest edge of this trend. Fraud and scam complaints made up 35 percent of all banking-related complaints received by the UK Financial Ombudsman Service in 2023 to 2024, a total of 27,675 complaints and a 28 percent rise on the previous year. In the first quarter of the 2024 to 2025 financial year, consumers lodged 8,734 fraud and scam complaints, and over half concerned authorised push payment scams. These numbers show a market where AI tools amplify existing fraud channels rather than inventing wholly new ones.

Emerging economies carry extra risk. Studies of EU converging economies such as Romania, Bulgaria, Poland, and Croatia find lower cybersecurity spending relative to GDP, heavier reliance on imported ICT services, and weaker capacity for rapid incident response. The result is a disproportionate exposure to AI-enabled cyber threats aimed at financial systems.

Why AI Attacks Create Legal Risk

The legal risk stems from a simple chain. A business holds personal data and provides services. An AI-enabled attack breaches that data or facilitates fraud. Affected customers then seek redress. They may sue for negligence, breach of contract, or a data protection failure. They may also complain to a regulator. Both routes carry financial penalties and reputational damage.

Liability already deters technology adoption. Research on EU firms found that 33 percent view liability for damage as the top external obstacle to AI adoption, and 29 percent cite the need for new laws. This concern is rational. When harm results from a faulty or exploited AI system, proving what happened is difficult. The internal logic of a model is often opaque, a problem described as the black box issue. Claimants must show causation, and that usually requires expert testimony and access to proprietary systems.

The evidence problem cuts both ways. AI-generated content, whether text or image, is often indistinguishable from human-generated content. There is no standardised reporting framework for AI-enabled fraud. This gap obscures the role an AI system played in a given loss and places a heavy burden of proof on the affected party. It also leaves defendant firms uncertain about the standard they must meet.

The EU Regulatory Framework

The European Union has built a layered set of rules. The General Data Protection Regulation governs personal data. Article 24 requires appropriate technical and organisational measures against threats to individuals. Article 82 gives individuals a right to compensation for both material and non-material damage caused by a data protection breach. A firm that suffers an AI-enabled breach can therefore face direct claims where its safeguards fall short.

The Artificial Intelligence Act adds a risk-based structure. It classifies general-purpose AI models and imposes duties that scale with risk. Providers of models that pose systemic risk must conduct model evaluations, including adversarial testing known as red teaming, assess and mitigate risks, report incidents, and maintain adequate cybersecurity. A model is presumed to pose systemic risk if its training involves more than 10 to the power of 25 floating-point operations. Obligations for general-purpose models apply from 12 months after the Act entered into force, with a 24-month transition for models already on the market.

Two liability instruments sit alongside the Act. The revised Product Liability Directive extends strict liability to all AI systems and AI-enabled goods, and it accepts that a system can become defective from knowledge learned after deployment. The proposed AI Liability Directive introduces fault-based procedures, including a rebuttable presumption of causality and disclosure duties that shift the burden of proof toward providers and deployers. The AI Liability Directive remains parked in the legislative process, so a gap persists between the two regimes.

Further instruments target security directly. The Cyber Resilience Act sets security requirements for digital products. The NIS2 Directive raises cybersecurity obligations across essential sectors. The Digital Operational Resilience Act, known as DORA, governs operational resilience in finance. Together these rules form a framework where a single AI-enabled incident can trigger obligations under several laws at once.

The UK Position

The United Kingdom takes a principles-based, sector-led approach. Its 2023 White Paper set out principles of safety, security, transparency, explainability, fairness, accountability, governance, contestability, and redress. Rather than a single AI statute, regulators such as the Financial Conduct Authority and the Information Commissioner’s Office issue guidance within their remits. This design favours innovation but offers less clear enforceability in high-risk contexts.

Data protection still provides the main legal footing. The Data Protection Act 2018 and the UK GDPR require lawful, fair, and transparent processing. They give individuals rights of access, rectification, erasure, and a right not to be subject to automated decisions under Article 22. Individuals can bring civil claims for damage, and the ICO can enforce compliance under sections 149 to 152 of the Act. Financial firms that use AI must meet these duties or face claims and penalties.

Recent changes signal a shift in institutional responsibility. As of July 2025, the UK reduced the statutory ceiling for fraud-related compensation payable by financial institutions to 85,000 pounds per incident, aligning with the Financial Services Compensation Scheme cap. This recalibration affects client expectations and litigation strategy, particularly in AI-fraud cases involving large losses. The FCA Dispute Resolution rules set out how firms can be held accountable and how consumers can seek redress when an AI system causes harm.

Liability Gaps and Litigation Challenges

The core difficulty in AI-fraud litigation is causation. A customer must show a direct link between an AI system failure and the loss. This often requires transaction records, communication logs, code reviews, and expert analysis. Where the system is proprietary and opaque, the claimant may struggle to obtain the evidence needed. The EU disclosure mechanisms aim to ease this, but they apply unevenly and depend on classifications under the AI Act.

Scope mismatches widen the gap. The Product Liability Directive requires evidence disclosure for all AI systems, while the AI Act mandates continuous logging only for high-risk or systemically risky systems. Standard general-purpose models face transparency and documentation duties but no blanket obligation to record events across their lifecycle. A claimant harmed by a standard model may therefore find little recorded evidence to support a claim.

For firms, the exposure is concrete. Non-compliance with GDPR, the Data Protection Act, or FCA principles can lead to civil claims and regulatory action. Penalties can be substantial, and reputational damage can cause loss of customer trust and business. Firms also face liability in contract and tort, alongside risks tied to data misuse, discrimination, bias, and anti-competitive conduct where AI decisions drive outcomes.

Practical Steps for Businesses

Firms can reduce legal exposure through disciplined governance. Map every AI system in use, including third-party tools, and classify each by risk. Document data sources, testing, and model evaluation results. This record supports compliance under the AI Act and provides the evidence trail that limits liability if an incident reaches court.

Strengthen technical safeguards against AI-enabled attacks. Deploy detection for deepfakes and synthetic content, enforce multi-factor verification for payments, and run adversarial testing on defensive systems. Maintain logging that captures interactions and timestamps, because clear records help both defence and regulatory reporting. Align incident response with NIS2, DORA, and national reporting duties so that notification is fast and complete.

Governance must reach the board. Assign clear ownership for AI risk, review duty of care against current guidance, and update contracts to allocate liability with vendors and customers. Train staff to recognise AI-driven social engineering, since human error remains the common entry point. These measures do not remove risk. They demonstrate reasonable care, and reasonable care is the standard against which courts and regulators will judge a firm.

Outlook

AI-powered cyberattacks are now a standing feature of the business environment. The legal response is still forming, with the EU building a layered framework and the UK relying on principles and sector guidance. Gaps in causation, disclosure, and classification leave real uncertainty. Firms that act now, by documenting systems, hardening defences, and clarifying liability, place themselves in the strongest position as the rules harden through 2026 and beyond.