A switch controlled elsewhere

When Christine Lagarde opened the tenth annual conference of the European Systemic Risk Board in Frankfurt on Thursday, she used an event from June as her example. A US export-control directive concerning two advanced AI models, she said, led their provider to suspend access. "For Europe, the result was an abrupt cut-off." Her speech's footnotes name the provider and the models: Anthropic's statement of 12 June on suspending access to Fable 5 and Mythos 5, which, they add, it did for all customers, including those in the United States. Access to the general-use model came back a few weeks later. The model with fewer cyber safeguards, Lagarde said, "remained limited to organisations vetted by the US administration."

Lagarde is president of the European Central Bank and chair of the ESRB, the body created after the financial crisis to watch for risks that build up across the whole system rather than in one firm. "That incident passed without any discernible disruption to the financial system," she said. Her point was about what happens later, if "almost every institution depends on a select few models or AI firms to adjust trading strategies or find vulnerabilities in their systems. In that scenario, a loss of access could prove highly disruptive to financial stability."

The speech set out three risks. In trading, AI agents "may pursue goals in ways their human overseers did not intend and cannot detect". In cyber resilience, "an attack on shared technology has the potential to disrupt several firms at once". And in geopolitics, tensions raise the threat of cyberattacks and can "prompt restrictions on the frontier models that firms use to defend themselves." Each is serious alone, she said, but "together they can interact and compound across the financial system."

The numbers she gave show where adoption stands. Nearly nine in ten significant euro area banks use generative AI. Seven in ten EU securities firms that responded to a recent survey expected to increase their spending on it. Autonomy is still rare: this year only 5 per cent of asset managers said they gave AI autonomous or semi-autonomous authority over investment recommendations or trades. She said that share "may well increase over time", and pointed to one of the world's largest hedge funds, which her footnote identifies as Bridgewater, launching a strategy in which AI is the main decision-maker, with the stated goal "to build a fully artificial investor that can outcompete humans".

On cyber, she cited the ESRB's warning from July and the speed of the change. In a simulated 32-step attack, models released at the end of 2025 completed about a third of the steps on average. The latest models completed every step. The ESRB has warned that the gap between a first exploit and widespread automated exploitation "could fall from weeks to hours". Because banks have to test a fix before putting it into essential systems, and attackers can use a weakness as soon as they find it, the board expects attackers to have the advantage "in the short to medium term". That makes frontier models more valuable to defenders. It also makes losing access to them more costly.

Her conclusion was about who owns the models. Frontier development is concentrated in the United States and China, she said, and that raises the question of "who controls access to the models its financial institutions may come to rely on? That question was once hypothetical. Since this summer, it no longer is." She compared the risk to the faulty software update that grounded flights and disrupted banking around the world in the summer of 2024. Europe, she argued, needs AI capabilities of its own and a place in the AI supply chain, "so that access to tools that are vital for its financial resilience does not depend on a switch controlled elsewhere."

She also offered a model for global rules: the Cold War agreements under which the US and the Soviet Union limited their nuclear capabilities while still competing. She acknowledged the difference. The United States leads in frontier AI and "may view any global agreement as a means to slow it down. Yet such a view would be shortsighted." The ESRB's immediate requests are closer to home: review the cyber defences around critical financial systems, and make sure firms plan timely responses and work together where an attack could spread across the sector. The example she chose for Europe's exposure was not a market crash or a bank failure. It was a supplier switching off a model.