Researchers say 44 AI agents drove attacks on Japanese targets

Cybersecurity researchers at TrendAI disclosed on 27 October 2026 that they had analysed three cyber attacks against organisations in Japan in which artificial intelligence played a central role. In each case, according to the researchers, AI was either in the driving seat or assisting a human operator. The system behind the activity is named CyberStrikeAI and uses 44 AI agents to automate hacking operations.

The disclosure is based on TrendAI's analysis of three exposed server environments used by the threat actors to conduct the operations. Exposed servers of this kind are machines left reachable from the internet without adequate access controls, which can allow outsiders to inspect the tooling and configuration an attacker has staged there. TrendAI said the three environments it examined were distinct from previously published cases.

The number of agents involved is notable because it points to a division of labour inside the attack tooling. An AI agent, in this context, is a software component that can take actions on its own, such as running commands or calling other tools, rather than simply generating text for a person to act on. TrendAI's account places CyberStrikeAI on a spectrum that runs from fully autonomous operation to human-directed use, with different attacks falling at different points along it.

The researchers did not identify the organisations that were targeted, and the available material does not say what data, if any, was accessed, how long the intrusions lasted, or when they began. No vendor product was named as the source of the compromise, and TrendAI did not publish an attribution to a known group in the material disclosed. The root of the problem, as described, is the misuse of legitimate automation tooling rather than a software flaw with a patch.

For defenders, the practical point is narrow but real. Attack tooling that spreads work across dozens of agents can run reconnaissance, access attempts and follow-on actions in parallel, and at a pace that does not depend on an operator being at a keyboard. Detection approaches that assume a single human working through a target list may fit that pattern poorly.

TrendAI's analysis is the only account of these three incidents currently available, and further detail, including any confirmation from the affected organisations in Japan, has not been published. Readers should treat the specifics, including the agent count and the degree of autonomy in each case, as the researchers' assessment of what they found on the exposed servers rather than as independently confirmed fact.