In April this year, the United States Federal Reserve joined the Treasury Department to convene a meeting of large bank executives to remind them that a new AI model could pose cybersecurity risks. However, CNBC reported that at least three months after the warning was issued, the Fed itself had not been given access to the model.
The model has been opened to some institutions
The model, Claude Mythos Preview, is part of Project Glasswing of the Anthony Network Security Project. Anthropic states that the model is good at identifying weaknesses and security gaps in the software and has been opened to some banks and institutions to detect and repair system deficiencies.
The Fed's still working on it.
It was reported that, as of 15 July, the Federal Reserve was still trying to qualify for Mythos and it was not clear whether it had since been granted. The Fed refused to comment, and Anthropic did not respond immediately.
At last week’s congressional hearing, Federal Reserve Chairman Kevin Warsh stated that the Fed did not have decision-making power over model access, but he had expressed his views within the government, hoping to gain more access to the Frontline AI model not only for the Fed, but also for other agencies, so that potential gaps could be identified and closed as soon as possible.
Anthropic Expand Project
In early April, when Claude Mythos Preview and Project Glasswing released, Anthropic stated that about 50 institutions were granted access. Companies named include Chase Morgan and technology companies such as Amazon, Apple and Google. Anthropic also stated at that time that there had been continuous communication with United States Government officials, including such departments as the Cybersecurity and Infrastructure Security Agency.
By June, Anthropic had expanded the scope of Project Glasswing to over 150 institutions in 15 countries. According to Daniel Newman, Chief Executive Officer of Futurum Group, a research institute, it is surprising that even the Fed, which is responsible for policy coordination of the financial system, was not included in the first assessments.
Export controls have disrupted the open rhythm.
The report mentions that the Mythos outreach process has been accompanied by a wave in recent months. In June, Anthropic indicated that, as a result of the Federal Government ' s export control directives on national security grounds, the company once closed Mythos 5 and the wider release version of Fable 5. Subsequently, Howard Lutnick, the United States Secretary of Commerce, authorized the company to reopen to some “credible partners” and the restrictions were subsequently completely lifted.
According to CNBC, the incident reflected the lack of effective coordination following the more active involvement of the Trump Government in AI. At the same time, Chinese model competition is rising. Kimi K3 released this month by Moonshot AI demonstrated better than OpenAI and Anthropic products in some industry benchmark tests, further increasing pressure on U.S. agencies to pursue forward models.