US backs Anthropic Mythos as AI edge over China

US Treasury Secretary Scott Bessent said Anthropic’s new Mythos AI model could help the US stay ahead of China.

Anthropic said Mythos is strong at finding software and system vulnerabilities and is being released to a small group of selected users.

Bessent’s remarks came days after he and Federal Reserve Chair Jerome Powell met Wall Street banks over concerns that Anthropic’s latest model could raise cyber risk.

Anthropic remains in a dispute with the Pentagon after it labeled Anthropic a US supply chain threat.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Anthropic’s Pentagon dispute ties to AI ethics and split court rulings

Tensions with the Pentagon started after Anthropic barred its AI from use in autonomous weapons or mass surveillance of American citizens 1.
The Pentagon told the court it needs freedom to use purchased technology for any lawful purpose 1.
A federal judge in San Francisco temporarily blocked enforcement of the Trump administration directive that ordered agencies to stop using Anthropic and its chatbot Claude 2.
A separate Washington, D.C., appeals court kept the Pentagon’s “supply chain risk” designation in place for now 3.

Mythos stays limited to select partners due to dual-use cybersecurity risks

Anthropic launched “Project Glasswing” using its unreleased frontier AI model, Claude Mythos Preview. It is limited to participating companies and a small set of vetted organizations, with no plan for a public release 4.
The partner list includes Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, plus Palo Alto Networks 4.
UK government testing reported by Gigazine found the model could autonomously run a full network takeover in a corporate network simulation. The agency also warned it could not say for sure the model can break into well-defended systems with active security tools and staff 5.
Across 10 trials, the model finished the simulated 32-stage attack three times. The test is estimated to take a human 20 hours 5.
Anthropic also reported Mythos identified thousands of zero-day vulnerabilities, previously unknown software flaws. It also found a 27-year-old bug in OpenBSD, an open-source operating system. The same skills could help attackers spot weak points and map attack paths 4.