Anthropic has introduced Mythos Preview, a specialized large language model designed for high-level cybersecurity tasks. The company claims the model has already identified thousands of critical vulnerabilities, including zero-day flaws in major operating systems and browsers, as well as a 27-year-old error in OpenBSD.
According to Anthropic, the model can generate functional exploits. The company reported that even users without formal security training were able to use the tool to achieve remote code execution on target machines.
A new era of strategic infrastructure
Because more than 99% of the identified vulnerabilities remain unpatched, the model's full capabilities cannot be publicly audited. Anthropic stated that "Mythos Preview is just the beginning," signaling a shift toward models that may not be released to the general public due to their potential social risks.
This shift suggests a new category of "frontier models" where the cost of widespread release outweighs commercial value. British regulators are currently evaluating the impact of such technology on critical infrastructure, and Anthropic has confirmed ongoing discussions with U.S. government officials.
Competitors are already moving to secure controlled access. OpenAI launched a "trusted access" scheme for its cybersecurity-focused models, utilizing identity verification and enterprise-only programs to manage risk.
The scale of investment in this sector is massive. Reuters reports that Alphabet, Amazon, Meta, and Microsoft could collectively invest $650 billion in AI infrastructure by 2026.
Access to the most advanced tools is currently concentrating among large cloud providers, systemic banks, and defense teams. Initiatives like Project Glasswing—which includes Amazon Web Services, Apple, Cisco, Google, JPMorganChase, Microsoft, NVIDIA, and the Linux Foundation—demonstrate this trend.
Anthropic has also expanded access to 40 additional organizations, providing up to $100 million in credits via enterprise platforms from Amazon, Google, and Microsoft. This leaves independent academics, small startups, and the general public without access to the most powerful defensive and offensive capabilities.