The government is rewriting security standards for next-generation artificial intelligence because AI is changing the very nature of cyberattacks. AI that once explained attack methods or wrote code is now carrying out multiple stages on its own — from scanning for vulnerabilities to breaching systems and stealing data — cutting the time and manpower an attack requires while widening the damage. Attackers can hit more targets at lower cost, while defenders have less and less time to detect and respond to threats.
Anthropic, OpenAI and Google are discussing a plan to jointly establish a standards body for the AI industry, U.S. technology news outlet The Information reported on the 13th. As concerns grow that AI could slip beyond human control and calls mount to slow the pace of development, efforts to build industry-level safeguards are taking shape.
The speed and scale of AI-assisted attacks are already evident in actual breaches. According to a report Anthropic released on the 10th of this month analyzing misuse of its AI model Claude between December of last year and August of this year, attackers believed to be affiliated with the international hacking group ShinyHunters broke into a software-as-a-service (SaaS) provider and used that access to extract data from about 200 client companies. They then stole login session data containing more than 2,100 bundles of authentication tokens (Azure AD tokens), which function as digital access badges, in roughly 34 hours. Anthropic said AI agents performed nearly all of that work. In another attack, a single stolen developer token was used to gain administrator privileges over a victim's cloud environment in about three hours. Cybercriminal groups steal data routinely, but as AI advances, the scale is growing sharply and the pace is accelerating.
Cases in which multiple AIs are deployed at once to divide up an attack have also been confirmed. One criminal group in China designed and built an autonomous workflow that let Claude search around the clock for vulnerabilities in targets and research exploit code to take advantage of them. The group targeted about 50 organizations worldwide, including companies in education, health care, finance and manufacturing as well as government agencies, and stole a substantial volume of data. Agentic AI played a central role here as well. In particular, the group used an "agent swarm" approach in which one AI agent divides up tasks and assigns them to multiple sub-agents, allowing several attack operations to run in parallel without step-by-step human direction.
"Sophisticated attacks no longer require sophisticated human attackers," Anthropic said. "Every phase of an attack — from reconnaissance and tool development to data processing and exploitation — has been advanced through AI." The company also warned that "AI's role is shifting toward increasingly autonomous forms," adding that "unless stronger security measures are taken as AI model capabilities continue to improve, the risks posed by AI will grow further."
While analyzing second-quarter breaches this year, the Google Threat Intelligence Group (GTIG) identified a case in which AI sharply compressed attack timelines. Working inside a cloud environment it had already compromised, the attacker used agentic AI to plan, build and execute a large-scale credential theft operation in under six hours. The AI autonomously handled tasks such as scanning for vulnerabilities and resolving errors during execution. GTIG concluded that defenders' response times are shrinking as attackers move beyond asking AI piecemeal questions to automating multiple stages of an attack.
In a survey of 602 companies that suffered hacking incidents over the year from March of last year through February of this year, IBM found that 25% said they had been targeted in attacks that used AI. That figure was up 56% from the previous year, and damage from AI-driven attacks added an average of $1 million (about 1.347 billion won) in extra costs per breach.
Cyber threats in Korea are also rising rapidly. Cyber breach incidents reported in the first half of this year totaled 1,236, up 19.5% from the same period last year, according to the Ministry of Science and ICT. "With the spread of generative AI, not every attack is fully automated yet, but when the time and cost of preparing an attack fall, attackers can analyze more targets in less time than before," the ministry said. "The misuse of AI is becoming more than a factor that amplifies the speed and scale of existing cyberattacks — it is emerging as an independent new type of cyber threat."
Experts said cybersecurity must therefore evolve from a focus on detection to a focus on verification and speed. "High-performance AI like Mithos is lowering the cost of finding vulnerabilities," said Yang Jong-heon, head of the integrated analysis division at S2W. "Security from here on should put more weight not on finding vulnerabilities but on how to verify the ones AI uncovers, in what order to fix them, and how to build defenses faster than attackers can."
