Bundesbank Warns Anthropic’s Mythos Could Expose Bank Software

Speaking in Rome on Tuesday, Bundesbank president Joachim Nagel warned that Anthropic’s Mythos model appears able to rapidly identify and exploit software vulnerabilities used by banks. Nagel, who also sits on the European Central Bank’s Governing Council, urged that European banks and supervisors be given quick access to the model so they can test its effects.

Nagel told the audience the model can help strengthen digital defenses and also be used by attackers. He warned that banks should not be expected to defend systems they cannot fully inspect. He added, “We must prevent the misuse of this technology,” and noted, “Early identification and mitigation of such risks are crucial for financial stability.”

Nagel described Mythos as capable of locating and exploiting security holes in software relied on by financial institutions. He called for regulators and affected firms across Europe to gain rapid access so they can test the model directly and assess both benefits and risks.

Nagel highlighted figures showing a gap in advanced AI capacity. He cited that U.S. institutions built 40 advanced models in 2024, China produced 15 and Europe produced three. Private AI investment in 2024 stood at about $109.1 billion in the U.S., $9.3 billion in China and $19.4 billion in Europe. He noted state AI investment estimates at $62 billion in China, roughly $3.3 billion in the U.S. and about $1.2 billion in the EU.

Anthropic this week announced an expanded multi-year agreement with Amazon Web Services to secure as much as 5 gigawatts of compute capacity for training and deploying its Claude family of models. Anthropic said new Trainium2 capacity will come online in the first half of the year and that nearly 1 gigawatt of combined Trainium2 and Trainium3 capacity is expected by the end of 2026. The company reported it currently uses more than one million Trainium2 chips and has launched Project Rainier, a large compute cluster.

Under the deal, Anthropic committed more than $100 billion over the next decade to AWS technologies, covering Graviton processors and Trainium chips from Trainium2 through Trainium4, with options for later generations. The agreement expands inference capacity in Asia and Europe and keeps AWS as the primary provider for mission-critical workloads. Anthropic reported that more than 100,000 customers run Claude on Amazon Bedrock. Amazon CEO Andy Jassy praised the custom silicon, saying, “Our custom AI silicon offers high performance at significantly lower cost for customers, which is why it’s in such hot demand.”

Nagel warned that concentration of advanced models in a small number of companies could leave European institutions at a disadvantage when assessing and defending against risks. Regulators and industry groups are focusing on how generative AI and autonomous agents could introduce new attack vectors, and he urged that banks and supervisors be permitted to inspect and test systems directly.

The figures Nagel cited underscore differences in AI capacity and financing across regions and, he noted, support the case for hands-on access by European firms and authorities so they can test, defend and regulate against potential misuse.