Anthropic has restricted public access to its latest artificial intelligence model, Claude Mythos, following internal testing that proved the system capable of automating cyberattacks. The company confirmed it will only provide access to a small group of partner organizations to prevent the technology from falling into the hands of malicious actors.
While Mythos was designed for general coding and complex problem-solving, its architecture allows it to analyze vast amounts of software code with speed and precision. During internal trials, the model successfully identified previously unknown vulnerabilities in widely used browsers and operating system components. In several instances, the AI even generated functional attack sequences to exploit those flaws.
Security risks and defensive strategies
The ability of Mythos to find bugs that have remained undetected for years has drawn significant attention from cybersecurity experts. According to reports from Wired, the model’s performance in identifying security gaps and mapping potential attack routes is comparable to that of professional human penetration testers.
Anthropic officials stated that the risks associated with an unrestricted release are too high. They warned that the technology could allow bad actors to compromise critical infrastructure or gain unauthorized access to sensitive systems. By keeping the model within a controlled environment, the company aims to monitor its behavior and prevent potential misuse.
To counter the inherent dangers of such powerful code-analysis tools, Anthropic launched Project Glasswing. This initiative focuses on the defensive application of the technology, encouraging participating organizations to use the AI to identify and patch security holes before they can be exploited by outsiders.
"The project seeks to coordinate efforts between companies to detect vulnerabilities before someone takes advantage of them for malicious purposes," the company noted in its technical report. Participants in this program are currently using Mythos to audit their own systems and report findings under strict supervision.
Anthropic has not provided a timeline for a broader release of the model. For now, the company continues to evaluate the software's performance while maintaining tight restrictions on who can interact with the system.
