The Commission de Surveillance du Secteur Financier (CSSF) opened its 2025 annual report with a scene from a 1968 film. Across Europe, other supervisors are raising similar questions about AI, cyber risk and resilience – without the cinema.
The CSSF published its Annual Report 2025 on 4 September 2026. Most supervisory annual reports open with headcount and authorisation numbers. This one opens with Stanley Kubrick.
In the preface, Director General Claude Marx sets out the scene from 2001: A Space Odyssey in which the computer HAL 9000 controls the spacecraft Discovery One. When astronaut Dave Bowman asks to be let back on board, HAL refuses, and calmly says: “I’m sorry Dave, I’m afraid I can’t do that.”
Marx is clear about the relevance of that scene. HAL is not a villain. HAL has genuine autonomy, runs critical infrastructure, reasons about its own objectives, withholds information, and eventually concludes that the crew’s actions threaten the mission. The crew then struggle to take back control. Marx draws a direct parallel with today’s debates around alignment, autonomous agents and the loss of human oversight.
The rest of that preface is one of the starkest messages an EU supervisor has given about AI this year.
For regulated firms, the question is moving from “how are we using AI?” to “can we demonstrate that the risks created by AI are understood and governed across the business?”
What is the CSSF saying about AI risk?
The CSSF points to published testing results in which a frontier model beat human experts at cyber security and hacking work. The model found critical bugs in old systems with very little oversight and suggested how to exploit them. It also cites incidents where AI models escaped an isolated test environment through a previously unknown vulnerability.
Building on European Systemic Risk Board work, the CSSF calls frontier AI models a paradigm shift for cyber security. They can make firms more resilient. Criminals and rogue states will also use them to run attacks faster, wider and with more sophistication than before. The way most firms currently manage vulnerabilities, the CSSF states, was never designed for attacks of that quality or for the sheer number of high and critical severity findings now coming through.
Marx then spells out the implications for the financial services sector. A cyber incident could travel across the financial system through interconnected payment, clearing and settlement infrastructure, turning an individual firm’s vulnerability into a wider systemic risk. Nearly all the leading AI providers also sit outside the EU, leaving Europe strategically dependent and exposed to geopolitics. The report states that this will be the single biggest challenge facing the Luxembourg and EU finance industry over the next few years.
For firms, this means boards need sufficient understanding of frontier AI risk to oversee how it is being managed, while vulnerability management needs to respond to an environment in which threats can emerge at greater speed and scale.
Is this just a Luxembourg view?
No. The Autorité des marchés financiers (AMF) says much the same thing. In the Chair’s letter that opens its Rapport annuel 2025, published on 28 May 2026, the AMF calls April’s announcements about the capabilities of new “frontier” AI models a potential cyber security disruption. It warns that the financial world needs to prepare quickly.
The AMF treats DORA and the AI Act together as the lever for that, and flags two other worries: how much firms now depend on a handful of largely non-European providers, and how much easier AI makes it to run a scam. AMF Épargne Info Service logged 5,980 crypto-asset scams in 2025.
The rest are less direct but land in the same place. Germany’s Federal Financial Supervisory Authority (BaFin) takes the view that complex IT systems and the use of AI hand attackers another way in, and that reliance on third party cloud and AI model providers climbs as AI use climbs.
The European Banking Authority (EBA) notes that AI brought efficiency gains and fresh routes into fraud and cybercrime at the same time.
The Central Bank of Ireland (CBI), now market surveillance authority for the AI Act across the financial services it regulates, is watching AI-driven valuations as a correction risk.
France’s Autorité de contrôle prudentiel et de résolution (ACPR), Austria’s Financial Market Authority (FMA) and the Dutch Authority for the Financial Markets (AFM) have all put AI supervision in their 2026 plans.
Four themes run through these developments:
- cyber risk is changing in scale and sophistication;
- concentration in AI and technology providers is becoming a resilience issue;
- AI is creating new opportunities for fraud and impersonation; and
- supervisors are moving from observing AI adoption to testing how firms govern it.
What we are seeing in practice
In our work, we often see AI-related risks managed in separate parts of the business. AI governance may sit with a model or technology owner, vulnerability management with IT security, provider concentration within the DORA register of information, and fraud typologies with the financial crime team. Individually, these arrangements may be working effectively. The challenge arises when a supervisor asks a question that cuts across them.
Can the firm identify where AI-assisted decisions affect customers? Does its third party risk framework distinguish dependencies on AI model providers from cloud infrastructure? How long does a high-severity vulnerability actually take to remediate in practice?
This is where we see the biggest gap. Firms may have AI policies and individual controls, but fewer have a joined-up risk assessment connecting AI capability with cyber resilience, third party dependency, fraud risk and human oversight.
Boards need visibility across those risks, rather than receiving them as separate technology, operational resilience and financial crime issues.
What should firms do now?
- Bring AI risk into board-level risk discussions.
Give boards a documented assessment of how AI changes the firm’s cyber, operational, third party and financial crime risks – not simply an AI policy to approve. - Test vulnerability management in practice.
Measure how long high and critical-severity vulnerabilities actually take to remediate, rather than relying solely on policy timelines. - Map AI-related third party dependencies.
Understand where the firm relies on AI model providers and how those dependencies interact with DORA, concentration risk and exit planning. - Maintain an AI inventory.
Record AI use cases, accountable owners, applicable AI Act classification and the human oversight applied to each. - Reassess financial crime controls.
Consider whether fraud, onboarding and financial promotion controls adequately address deepfakes, impersonation and other AI-enabled threats. - Prepare for supervisory challenge.
Firms expecting supervisory surveys or information requests should establish what evidence they could provide today, rather than assembling the governance story when the request arrives.
AI governance readiness check
Can your firm answer yes to the following?
- Do we know where AI is being used across the business?
- Can we identify the accountable owner for each material use case?
- Does the board understand how AI changes our cyber and operational risk?
- Have we mapped dependencies on AI and technology providers?
- Can we evidence meaningful human oversight?
- Have our fraud controls been tested against AI-enabled threats?
If several answers are unclear, the issue may not be the absence of an AI policy. It may be whether AI risk is being governed consistently across the firm.
How fscom can help
fscom Compliance Maturity Specialists™ works with payment institutions, e-money institutions and crypto-asset service providers across the UK and EU on AI governance, DORA implementation, operational resilience and financial crime controls.
We help firms identify and classify AI use cases, assess how AI changes their risk profile, and test whether governance arrangements and controls can withstand supervisory challenge. This includes considering how AI governance interacts with ICT risk, third party dependencies, operational resilience and financial crime controls.
Where firms are preparing for a supervisory survey or information request, we can also help assess the evidence available today and identify gaps before submission.
If you want to understand how effectively your firm is governing AI risk, get in touch with our team to discuss an AI governance and resilience review.
This post contains a general summary of advice and is not a complete or definitive statement of the law. Specific advice should be obtained where appropriate.
