U.S. authorities have launched an emergency mobilization, convening top Wall Street bankers amid growing concerns that new, more powerful artificial intelligence models may amplify cyber risks.

US Treasury Secretary Scott Bessent and Fed Chairman Jerome Powell convened an emergency meeting on 7 April in Washington, D.C., calling on major banks to assess the risks posed by new AI tools.

The focus was on Anthropic’s new model, known as Mythos, which reportedly has the ability to locate and exploit vulnerabilities in operating systems and browsers if properly directed by the user.

The meeting, which was not previously made public, underscores that Washington sees a new generation of cyberattacks as one of the most significant risks to the financial system.

The banks involved are described as “systemically important,” meaning their stability is critical to the global financial system.

The authorities have called for ensuring that these institutions take adequate measures to protect their infrastructures against potential attacks that leverage advanced AI models.

The Mythos model and fears of abuse

Anthropic itself has shown caution, limiting the availability of the Mythos model to a limited number of large technology and financial services companies.

The system, introduced as the Claude Mythos Preview, is considered by the company to be so powerful that a widespread release is not planned at this time.

The goal is to be used primarily to identify and fix vulnerabilities in critical software before they can be exploited maliciously.

As part of this, a consortium of more than 40 companies, known as Project Glasswing, is being formed, involving giants such as Apple, Amazon, Microsoft and Google, as well as organizations such as the Linux Foundation.

Anthropic is committing up to $100 million in credits to use the model to enhance cybersecurity.

Meta in the superintelligence race with Muse Spark

Going on the offensive in the AI race is Meta, introducing its new Muse Spark model, the product of a costly investment to develop “superintelligence” i.e. AI that can outperform human cognitive abilities.

The Muse Spark is the first model in a new series, known internally as Avocado.

Unlike previous Llama models, Meta is not going full open sourcing, but is initially only offering limited access (“private preview”) to selected partners.

The model will initially be available through the Meta AI app and will gradually be integrated into platforms such as WhatsApp, Instagram, Facebook and the company’s smart glasses.

According to Meta, Muse Spark is designed to be “small and fast” but capable of handling complex queries in areas such as science, math and health.

Independent evaluations show that the model approaches the performance of leading competitors in language and visual understanding, but lags in programming and abstract thinking. It was ranked fourth in an overall Artificial Analysis evaluation index.

The company’s CEO Mark Zuckerberg has already foreshadowed a gradual improvement in performance, noting that the first models will mainly show Meta’s “rapid evolutionary trajectory.”