ECB Summons Banks to Patch Vulnerabilities Discovered by Anthropic’s Claude Mythos

ECB’s Three-Pronged Crisis Response

On May 26, 2026, the ECB convened a special meeting with banks to address the cybersecurity risks revealed by Claude Mythos Preview and similar advanced AI models . The meeting crystallized a regulatory approach built around three immediate demands.

Patch faster, with no excuses. ECB Vice Chair Frank Elderson warned banks that "the clock is ticking" and told them to "significantly accelerate" their security efforts to fix flaws the model had exposed . He described it as urgent work to "fix flaws that have been exposed" .

Share what you know. Because European banks are largely excluded from the limited consortium testing the model, the ECB asked U.S. banks that do have access to share insights with their European peers. Elderson acknowledged the situation was "unfortunate" but emphasized that "lack of access to the model is no excuse for inaction" .

Answer the supervisor. The ECB is using its regular supervisory dialogue to systematically quiz banks on their preparedness for AI-driven cyber threats, gathering information on threat exposure, patch cadence, and defensive tooling . The approach mirrors coordinated global regulatory action, though the ECB has not issued the kind of high-profile summons seen from the U.S. Treasury .

The urgency is backed by real-world validation. The UK’s AI Security Institute found Mythos Preview cleared 73% of expert-level Capture the Flag challenges, a benchmark no AI model could pass before April 2025. Mozilla shipped Firefox 150 with 271 patches for vulnerabilities found by the model .

What Claude Mythos Preview Actually Found

The vulnerabilities are not theoretical. Within a controlled evaluation period, the model autonomously identified and produced working exploits for thousands of high-severity flaws, including a 17-year-old remote code execution flaw in FreeBSD’s NFS server that grants unauthenticated root access to any internet-connected attacker, and a 27-year-old crash vulnerability in OpenBSD that had evaded human auditors for decades .

The scope is systemic rather than vendor-specific. Every major operating system and web browser is affected . Over 99% of the vulnerabilities discovered have not yet been patched, leaving a vast attack surface exposed across the financial sector’s legacy technology infrastructure .

The model’s exploitation capability goes beyond passive detection. It autonomously demonstrated exploitation paths for critical flaws in widely deployed software, confirming findings with working proofs of concept during the evaluation window . Cybersecurity experts now view the model as posing significant challenges to the banking industry and its legacy technology systems .

Why Europe Is Flying Blind

The most acute challenge facing European institutions is not just the existence of the vulnerabilities but the asymmetric access to the tool that found them. Anthropic structured the release of Claude Mythos Preview through Project Glasswing, a gated research preview with access prioritized for defensive cybersecurity use cases—but limited to a consortium of primarily U.S. tech partners that includes AWS, Google Cloud, and CrowdStrike .

European banks, regulators, and cybersecurity firms are largely locked out. They must remediate flaws they cannot independently probe or verify with the same AI capability . This creates a defensive asymmetry where attackers with access could exploit vulnerabilities that defenders in Europe cannot easily analyze or reproduce.

The regulatory gap is particularly frustrating for EU policymakers. The ECB and European Parliament are pressing for answers and action but lack the direct technological access that U.S.-based consortium members enjoy, complicating oversight and independent risk assessment . The Association of German Banks told S&P Global Market Intelligence that it is in "continuous dialogue" with its member banks, the Federal Ministry of Finance, BaFin, the German central bank, and European and international institutions—but that engagement remains reactive rather than proactive .

Anthropic has committed to reporting publicly within 90 days on the findings from Project Glasswing, a disclosure the industry is watching closely . Until then, European regulators are caught between an urgent operational threat and a tool they cannot touch, asking banks to patch faster against an adversary they can’t yet fully see.