In a move that highlights the growing tension between AI capability and AI safety, Anthropic has developed what it calls its most powerful cybersecurity AI model??nd then decided not to release it to the public. The model, internally referred to as Claude Mythos Preview (or Project Glasswing internally), represents a significant leap in AI-driven vulnerability research, but the company says the risks of immediate public release far outweigh the benefits.
The model was developed under a coordinated vulnerability disclosure program in partnership with twelve major technology companies: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. This impressive roster of partners reflects the seriousness with which the industry views AI-enabled cyber threats.
What makes Project Glasswing remarkable is not just its capabilities, but what it found. In testing, Claude Mythos Preview discovered a 27-year-old vulnerability in OpenBSD??ne of the most security-conscious operating systems in the world. It also found a 16-year-old flaw in FFmpeg, the widely-used multimedia framework, and multiple previously unknown vulnerabilities in the Linux kernel. These discoveries underscore both the model’s analytical power and the vast, largely unexplored landscape of security flaws lurking in critical infrastructure.
Rather than releasing the model publicly, Anthropic is taking a controlled approach. The company has committed million in API usage credits to partner organizations for legitimate security research, along with million in grants to open-source security organizations. This “white hat” approach allows the technology to be used for defensive purposes while preventing it from falling into the wrong hands.
Anthropic’s decision comes at a time when the company’s commercial position is strengthening significantly. The AI safety-focused company recently reported billion in revenue run rate?? staggering figure that represents doubled growth in just two months. The number of customers paying over million annually has also doubled, indicating that enterprises are increasingly willing to invest heavily in advanced AI systems.
The ethical implications of powerful AI in cybersecurity are profound. On one hand, AI systems like Project Glasswing have the potential to identify and patch vulnerabilities before malicious actors can exploit them. On the other hand, if such a system were to be weaponized or fall into wrong hands, it could be used to discover and exploit vulnerabilities at an unprecedented scale and speed.
Anthropic’s coordinated disclosure model offers a potential template for the industry. By partnering with defenders rather than releasing capabilities freely, the company attempts to balance the dual-use nature of AI security tools. Whether this approach becomes an industry standard will depend on how effectively it prevents both malicious use and the proliferation of similar capabilities by other actors.
The broader question remains: as AI capabilities continue to advance, how do we ensure that safety measures keep pace? Anthropic’s decision not to release Project Glasswing publicly is a notable data point?? company choosing restraint when it could have chosen expansion. In the high-stakes world of AI development, that choice may prove as significant as the technology itself.