Claude’s new model can independently crack the defense of entire corporations

New artificial intelligence models are increasingly confident in handling tasks that used to require hours of manual work by specialists. The latest test has shown that the development is faster than even industry observers expected — and we are no longer talking about theory, but about quite practical attack scenarios.

AI Security Institute conducted a series of tests of the Claude Mythos Preview model by Anthropic and recorded a significant increase in its cybersecurity capabilities. The results showed that the model confidently outperforms previous solutions and demonstrates progress in both standard tests and complex attack simulations.

Over the past two years, the approach to evaluating such systems has changed dramatically. While in 2023, models could barely cope with basic tasks, now Claude Mythos Preview is able to find vulnerabilities and exploit them on its own. During testing, the model was given access to the network and instructions, after which it was able to conduct multi-stage attacks that would have taken a human days to complete.

The experts paid special attention to Capture the Flag tasks, where it is necessary to find weaknesses and extract hidden data. At the complex level, which remained unattainable for AI until the spring of 2025, the new model successfully coped in 73 percent of cases.

The experiment with the simulation of an attack on a corporate network called The Last Ones was more revealing. The scenario includes 32 consecutive steps, from reconnaissance to complete infrastructure takeover. According to the authors of the test, a person would need about 20 hours to complete such a task. Claude Mythos Preview managed to complete the scenario in three out of ten attempts, and completed 22 steps on average. The closest competitor showed a much more modest result.

At the same time, the model has not yet coped with another scenario related to industrial systems, getting stuck at the stage related to classic IT infrastructure. This result does not allow us to draw unambiguous conclusions about its capabilities in an industrial environment.

The authors of the tests emphasize that the test conditions differed from the real world. The simulations lacked active defense mechanisms, detection and response systems, and consequences for suspicious actions. Therefore, it is too early to talk about the model’s ability to attack protected systems.

However, it is already clear that such tools can work effectively against weakly protected networks. Against this backdrop, experts urge companies to pay more attention to basic protection: installing updates in a timely manner, controlling access, and monitoring system events.

The development of such models will continue, and their protective potential is growing along with the risks. In the upcoming research, the team plans to test AI in conditions as close to reality as possible, including active countermeasures and monitoring.