Within two days, Anthropic, a San Francisco-based AI company, unveiled its new AI model focused on cybersecurity, named Claude Mythos Preview, sparking widespread discussion in the technology sector. The firm describes it as the leading model for cybersecurity applications, particularly in detecting previously unknown weaknesses in software code. Anthropic reports that the model identified thousands of critical vulnerabilities in major operating systems and web browsers, raising significant security concerns. To mitigate risks, the company has restricted its availability, noting the model’s potential to compromise systems.
The announcement has prompted a mix of reactions globally, with some experts expressing worry about its power and others questioning the extent of its abilities. Here is an overview of Anthropic’s latest development.
Claude Mythos is presented as a versatile AI model with strong capabilities in programming and logical analysis. These features enable it to examine extensive code repositories and uncover security issues that human specialists might overlook. Anthropic states that the model has detected flaws in systems that remained hidden for many years.
At present, access is limited to participants in Anthropic’s Project Glasswing. This group includes prominent organizations such as Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. Virtually all leading technology companies are collaborating with Anthropic to utilize the model.
Usage is confined to cybersecurity enhancements, allowing these partners to strengthen their protections without enabling offensive actions against others. Anthropic is also providing access to about 40 entities involved in essential software infrastructure, along with usage credits valued at $100 million (approximately Rs. 923.6 crore).
Regarding performance, Anthropic released a system card detailing evaluations on various benchmarks. The model surpasses the company’s previous top model, Claude Opus 4.6, in all pertinent assessments and excels particularly in cybersecurity metrics.
It achieved top scores on the USA Mathematical Olympiad, BrowseComp, and SWE-bench benchmarks. Notably, it earned a perfect 100 percent on Cybench, which evaluates proficiency in cybersecurity operations, making it the first AI to do so. Some tasks involve identifying and exploiting vulnerabilities in actual software. The model also dominated the CyberGym benchmark compared to other large language models.
The system card indicates that Mythos identified new vulnerabilities in Firefox, a widely used browser. Anthropic responsibly reported these to the developers.
The model remains unavailable to the general public. Logan Graham, who leads Anthropic’s Frontier Red Team, explained in an interview with NBC News that the AI can not only detect unknown vulnerabilities but also develop methods to exploit them. Its capacity to process large codebases allows it to execute sophisticated hacking operations.
In the wrong hands, this could lead to substantial harm. Through internal testing, Anthropic determined that these risks justify withholding public release.
Nevertheless, the creation of such an advanced tool prompts questions about oversight. As a profit-driven enterprise focused on competitive advantages, Anthropic may face calls for greater transparency regarding the model’s management and monitoring in the future.
