Anthropic is restricting public access to its new Claude Mythos Preview model after the AI demonstrated unprecedented cybersecurity capabilities, finding vulnerabilities across every major operating system and web browser. The company announced Project Glasswing—a consortium with AWS, Apple, Microsoft, NVIDIA, and major banks—to provide controlled access for testing while keeping the model away from public release. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened emergency meetings with CEOs from Goldman Sachs, Morgan Stanley, and other major banks to discuss the cyber risks.
This marks a significant shift in AI safety discussions from theoretical alignment concerns to immediate, measurable security threats. Unlike previous "too dangerous to release" claims around models like GPT-2, Mythos has already demonstrated concrete capabilities by identifying thousands of high-severity vulnerabilities. Goldman Sachs CEO David Solomon confirmed his bank has access and is working directly with Anthropic on security measures—a notable partnership between AI companies and the institutions most vulnerable to the technology's offensive capabilities.
Security researchers are particularly concerned about legacy banking infrastructure. Costin Raiu told Reuters that IBM-built systems from decades past would be especially vulnerable, calling them "ancient technologies powering the financial industry." The UK's AI Security Institute issued its own warning about Mythos, while the US government faces the awkward position of needing cooperation from a company it recently labeled a supply-chain risk over defense department usage disputes.
For developers, this creates a new precedent: advanced AI capabilities may increasingly come with restricted access and institutional gatekeepers. The banking industry's early access suggests future AI security tools will likely flow through established enterprise channels rather than open APIs or consumer products.