Anthropic’s latest AI model is restricted to use by only 40 institutions due to safety risks, as UK regulators rush to assess it

Gate News message: On April 12, the UK’s financial regulator is urgently assessing the potential risks of Anthropic’s latest AI model. On April 7, Anthropic announced the launch of its most powerful AI model, Claude Mythos Preview; however, during internal testing, the model demonstrated extremely strong cybersecurity capabilities: it successfully identified thousands of high-severity vulnerabilities and could carry out moderately complex hacking activities (including unauthorized access to the public internet, sending emails, and publicly publishing information). Based on security considerations, Anthropic decided not to publicly release the model, but instead to provide limited access through its “Project Glasswing” initiative to only 40 founding partner organizations (including Apple, Google, Microsoft, Amazon, JPMorgan Chase, and others). This move triggered a chain reaction in the market: cybersecurity-related stocks fell; U.S. Treasury Secretary Scott Bessent quickly convened Wall Street leaders to discuss risks to the financial system; and the Bank of England is also preparing to hold a meeting with banks and insurance companies to assess potential threats. At present, experts are divided on the matter: some believe the security concerns are genuinely real, while others question whether this action partly serves public-relations purposes, and they call for independent democratic institutions—not private companies—to decide the model release standards.