AI Cyberattacks Will Transform Industry Risk and Security

A new study from Carnegie Mellon University, in collaboration with Anthropic, has demonstrated that large language models (LLMs) can autonomously execute complex AI cyberattacks.

Efosa Udinmwen reports in TechRadar that by replicating the conditions of the 2017 Equifax breach, researchers demonstrated that AI systems can plan and execute attacks without any human intervention.

The results have significant implications for industries that rely heavily on data integrity and digital infrastructure.

The background to this research reflects the evolving use of LLMs in cybersecurity. Traditionally, AI tools were limited to controlled “capture-the-flag” exercises or code-level assistance.

The Carnegie Mellon team advanced this work by structuring LLMs as planners within a hierarchy of agents.

That allowed the AI to delegate tasks, adapt to its environment, and perform in conditions that more closely resembled real enterprise networks.

Under these controlled settings, the model not only deployed malware but also extracted data, exposing its potential for high-level operational autonomy.

The study revealed that the AI’s strength was not in raw coding, where models often falter, but in orchestrating tasks at a strategic level while offloading execution to sub-agents.

Although the experiments were conducted in a lab environment, and researchers emphasized the prototypical nature of the work, the system demonstrated an ability to replicate a major real-world breach without direct operator involvement.

Researchers noted that the same methodology could also be harnessed defensively, enabling AI to detect or block attacks.

Counsel advising technology-driven sectors should anticipate regulatory scrutiny over autonomous AI tools. Consider using contract terms that track evolving standards governing the dual-use potential of LLMs, in both AI cyberattacks and defense.