A recent study from Carnegie Mellon University and Anthropic, reported by TechRadar, reveals that large language models (LLMs) are edging into a new phase. Given the right architecture, they can plan and execute cyberattacks with minimal human oversight. The researchers built a multi-agent system with one “planner” LLM defining strategy and subordinate agents executing elements like malware deployment and data extraction. In a test modeled after the 2017 Equifax breach, this agent-based system identified vulnerabilities, exploited them, and carried out exfiltration without someone guiding each move by hand.

This is not yet a story of wild, uncontained AI attacks roaming freely online, but it is close enough to demand serious attention. If threat actors adopt similar approaches, what once required teams of hackers could be largely automated. That means faster attacks, more frequent breaches, and potentially new kinds of damage. At the same time, there is opportunity. Defenders can use similar AI setups for red teaming, vulnerability scanning, or simulations to find weak spots before they are exploited.

For those responsible for security, whether companies, institutions, or individual users, this study should be a wake-up call. It is time to rethink how threats are modeled, move beyond trust in static defenses, and assume adversaries may already be using agentic tools. Patch aggressively, limit privileges, invest in anomaly and behavior-based detection, and keep humans in the loop for critical decisions. Because when AI can plan itself, the side who prepares first sets the terms.