[DigitalToday reporter Chi-gyu Hwang] Anthropic warned in a recently leaked blog draft that new AI models under development, including its Claude "Mythos", could find and attack security vulnerabilities at a speed that cannot be compared with existing approaches.

Anthropic said, "Mythos foreshadows a wave of models that exploit vulnerabilities in ways that far outpace the speed at which defenders respond."

Anthropic is allowing some institutions to pre-test Mythos so they can strengthen their systems, and it is also privately warning government officials about the possibility of large-scale cyberattacks, Axios reported.

Shlomo Kramer (슐로모 크레이머), founder and CEO of Cato Networks, said, "Agent attackers are coming," calling it "a turning point in cybersecurity history."

He said a single AI agent can find and exploit vulnerabilities faster and more persistently than hundreds of human hackers. He added, "Behind Mythos are next-generation OpenAI models and next-generation Google Gemini, and in a few months Chinese open-source models will follow."

Evan Pena (에반 페냐), chief offensive security officer at security firm Armadin, said, "AI makes exploitation possible the moment vulnerabilities are discovered," but added that "AI models are weaker than human hackers at understanding context, such as identifying what information is most valuable in an organisation."

AWS' security research team said cases of AI-based cyberattacks are also continuing. In January, a Russian-speaking cybercriminal used multiple AI tools to hack more than 600 devices running popular firewall software in more than 55 countries.

The hacker used Anthropic's Claude and a model from Chinese AI startup DeepSeek. Bloomberg News reported that in February another hacker used Claude to carry out a series of attacks on Mexican government agencies, stealing tax and voter information.

Joe Lini (조 린), co-founder and CEO of Twenty, a company that sells offensive cyber capabilities to the U.S. government, said, "Even if machines handle execution, humans must always be responsible for the results," stressing the need to build weapons systems in which humans control decisions and outcomes.