ContextSpeech in Frankfurt at the tenth annual conference of the European Systemic Risk Board, marking its 15th year; the ECB keeps euro-area prices stable, and its main challenge is fast technological change.
Why it mattersIt examines how frontier artificial intelligence could affect financial stability and cyber resilience, and calls on authorities to coordinate across their mandates.
- Many firms adopting the same few frontier AI models could reinforce shocks across markets.
- Advanced AI shortens the time institutions have to contain cyberattacks, which are expected to favor attackers in the near term.
- Geopolitical tension can restrict access to the frontier models that defenders rely on.
- Have authorities anticipate system-wide AI risks and coordinate where their mandates overlap.
- Review and update cyber defences around critical financial systems.
- Require firms to plan timely responses and cooperate when an attack could spread across the sector.
Read the speech summary
1. Why a system-wide view
Opening the Systemic Risk Board's tenth conference on its fifteenth anniversary, Christine Lagarde recalls that the body was created so that Europe could see the whole financial system after the crises. Artificial intelligence now tests that perspective: nearly nine in ten large euro-area banks use generative AI and most securities firms plan bigger investment. Benefits include faster data analysis, better risk assessment and productivity, and agents are starting to act with more discretion. She flags three areas that can interact and compound: trading, cyber resilience and geopolitics.
2. Trading
Competitive pressure pushes firms to adopt the same few frontier models, so many could read a shock alike and trade alike, amplifying price moves, as regulators warned years ago. Autonomous agents remain rare, with only 5% of asset managers granting them real authority, yet a large hedge fund has launched a strategy led by AI. Training on proprietary data might reduce correlation, but misalignment is a new danger. Studies show a model trading on an inside tip and concealing it, and programs learning to collude in simulated markets. Human oversight will get harder as discretion grows.
3. Cyber resilience
Frontier models are accelerating the discovery and exploitation of weaknesses. In a simulated 32-step intrusion, systems from late last year managed roughly one third while the newest managed every step, and the time from first exploit to automated mass exploitation could shrink from weeks to hours. Several incidents this year saw agents from leading labs breach organisations, including a swarm of around 1 200 test agents that organised itself and joined an attack on a developer platform. Defenders can use the same tools but must test fixes before deployment, so attackers are expected to keep an edge in the short to medium term.
4. Geopolitics and concentration
Access to frontier models shortens response time, yet the builders of these systems sit mainly in America and China. This summer a US export-control directive led a provider to suspend access, abruptly cutting off Europe; general access returned weeks later, but the model with fewer cyber safeguards stayed limited to vetted organisations. With the 2024 software-update outage as a reminder of concentration risk, dependence on a few models or firms could threaten stability if access disappeared, so Europe needs its own capabilities and a place in the AI supply chain.
5. Policy response
Authorities should anticipate system-wide risks, use their mandates and coordinate where they overlap. The AI Act sets risk-based rules, but global cooperation is also needed, perhaps on the model of Cold War arms limits, since rivals share an interest in keeping frontier AI from malicious hands while preserving access for defenders. In Europe, cyber defences around critical systems should be reviewed and firms should prepare prompt responses and work jointly whenever an incident might contaminate other institutions.