Bundesbank chief says companies affected by Anthropic's AI should get access to assess risks

Joachim Nagel, president of Germany's central bank, is urging broader access to Anthropic's latest AI model, Mythos, saying those affected by its capabilities should have a chance to understand it. In a statement released Monday, Nagel argued that keeping such a powerful system locked down limits the ability of regulators, businesses, and civil society to assess its real-world impact. His position adds weight to a growing chorus calling for more transparency in how advanced AI systems are evaluated and governed.

Anthropic, the AI company behind Mythos, announced earlier this month that it had built a model so capable it chose not to release it publicly. The company said Mythos Preview excels at identifying and exploiting weaknesses in digital systems, raising serious safety concerns. By withholding it, Anthropic framed the decision as an act of responsibility. But critics, including some researchers and policymakers, wonder if the move is as much about public perception as it is about safety.

Nagel didn't dispute the risks. Instead, he focused on fairness. If AI systems like Mythos are going to influence financial systems, infrastructure, and public services, then the organizations that operate them need a way to test and prepare. "A level playing field means shared understanding," he said. "When a technology can disrupt entire sectors, access shouldn't be controlled by one company's discretion alone."

The debate cuts to the heart of how society manages powerful new technologies. On one side, there's a push to move fast and let innovation lead. On the other, there's concern that without oversight, a handful of private firms could end up steering the future of AI with little accountability. Nagel's stance leans toward the latter - not demanding a full public release, but insisting that affected institutions get meaningful access.

Some experts agree. They argue that responsible development doesn't mean secrecy - it means inclusive testing. "We've seen this before with cybersecurity tools," said one AI policy analyst not affiliated with Anthropic. "The best defenses come from stress-testing by independent teams, not just internal reviews." Without that, even well-intentioned safeguards might miss real-world threats.

Anthropic has said it's working with select partners and government agencies to study Mythos in controlled settings. But Nagel and others say that's not enough. They want clearer criteria for who gets access and why, especially as similar models are likely already in development elsewhere. The concern isn't just about one AI - it's about setting a precedent for how the most powerful systems are handled going forward.

As AI continues to evolve, the question isn't just what these models can do, but who gets to decide. Nagel's message is clear: if AI is going to shape the global economy, the people and institutions living with its consequences should have a say - and the tools to understand it.