The Fed’s AI Alarm: Anthropic’s Mythos Unattended

In April, the Federal Reserve and the Treasury Department convened an emergency meeting with the chief executives of the nation’s leading financial institutions. The primary concern: an advanced artificial intelligence model that could pose an unprecedented cybersecurity threat to the U.S. financial system.

The AI model in question, Claude Mythos Preview, developed by Anthropic, has demonstrated a remarkable capability for identifying software vulnerabilities. Anthropic launched it to a select group of banks and other organizations as part of a cybersecurity initiative known as Project Glasswing. However, a significant development emerged: for at least three months following its release, the Federal Reserve itself was denied access to Mythos. This left the institution, arguably the most systemically important in the global financial landscape, potentially exposed, even as other entities began to fortify their defenses.

As of July 15th, the Fed was still actively seeking access to Mythos, and it remains unclear whether the central bank has since secured it. Anthropic did not immediately respond to a request for comment, and the Federal Reserve declined to comment for this report.

The April meeting took place under the leadership of then-Fed Chair Jerome Powell, who, along with Treasury Secretary Scott Bessent, brought together the bank CEOs. More recently, in a congressional testimony that garnered little attention, Powell’s successor, Chairman Kevin Warsh, informed the Senate that efforts were still underway to gain access to Mythos and other cutting-edge AI models.

“We are not the arbiters of who gains access, but I have been vocal in expressing my concerns to authorities across the government regarding these vulnerabilities,” Warsh stated in response to questions from Senator Jack Reed. “I have been requesting access not only for the Federal Reserve but also for other institutions to a broad spectrum of these new artificial intelligence models, enabling them to enhance their security postures.”

Warsh further clarified that the need for access extended beyond just Mythos. “I wouldn’t want to solely focus on Mythos,” he added. “As these new models become more widely adopted, our banking system, and indeed the Federal Reserve, must make every effort to address any vulnerabilities we may have.” Warsh has, in his tenure at the Fed, expressed enthusiasm for AI, characterizing it as a transformative technology. However, the efficacy of addressing vulnerabilities without direct access to such advanced tools remains a pertinent question.

Anthropic unveiled Claude Mythos Preview and Project Glasswing in early April. The company indicated that approximately 50 organizations had access to the model at that time, naming a select few, including financial giant JPMorgan Chase and technology leaders such as Amazon, Apple, and Google. Anthropic also reported engaging in “ongoing discussions with U.S. government officials” concerning the model, including engagement with the Cybersecurity and Infrastructure Security Agency and the Center for AI Standards and Innovation.

In June, Anthropic expanded access to Project Glasswing, incorporating over 150 additional organizations across 15 countries. Daniel Newman, CEO of research firm the Futurum Group, expressed surprise at the Fed’s exclusion. “One would expect that the primary financial regulator, responsible for setting policy for the broader financial sector, would be at the forefront of evaluating new technologies,” Newman commented. “It’s certainly a surprising omission.”

The rollout of Mythos has generated significant interest and some confusion. Complicating matters was Anthropic’s complex relationship with the federal government, which led to export control directives. In June, Anthropic announced it had to disable access to Mythos 5 and Fable 5, an updated version of the model, in order to comply with a directive citing “national security authorities.” Commerce Secretary Howard Lutnick later granted the company permission to restore access to Mythos to a select group of “trusted partners.” The export controls were subsequently lifted entirely.

The current administration has significantly increased its focus on AI regulation following an executive order signed in June. However, the precise locus of AI policy decision-making remains somewhat opaque, and the situation at the Federal Reserve, now led by Warsh, is seen by some as an indicator of ongoing flux.

Adding to the dynamic, Chinese AI models are increasingly competing with leading offerings from U.S. companies like OpenAI and Anthropic. Moonshot AI, a Chinese startup, recently released Kimi K3, a model that has surpassed competitors on certain industry benchmarks. This development has fueled concerns among tech executives and government officials regarding the sustainability of U.S. leadership in the AI race.

David Sacks, a venture capitalist and former White House AI and crypto czar, voiced his apprehension, stating that Kimi K3’s performance is “concerning” and that “America is tying itself in knots.” He warned, “This is how you lose the AI race. The rest of the world won’t play by our rules if we bog ourselves down.”

Futurum’s Newman anticipates that the Fed will face a “catch-up” scenario, given its limited access to cutting-edge models compared to other agencies. “Every day, every week, whether it’s driven by China’s innovation or U.S. innovation, technology leaders within large institutions are confronting a constant barrage of new developments,” he observed. “The pace of advancement necessitates proactive engagement and, in the case of critical infrastructure like the financial system, a clear understanding of the risks and opportunities presented by these powerful new tools.”