TrendAI identifies three different uses of AI in cyberattacks
TrendAI researchers found different roles for AI in three exposed attack environments. One used CyberStrikeAI to automate work, while others relied on human operators with AI assistance. The findings distinguish target-validation lists from confirmed victims. The underlying models were not identified, and investigators did not establish that the three environments belonged to the same operator.
Artificial Intelligence··Night
Exposed servers reveal different workflows
TrendAI examined three separate attack environments targeting organizations in Japan in an investigation published on October 27. Researchers reconstructed activity from files, logs and tools left on exposed attacker servers. The Hacker News also reported the findings. In one environment, an open-source framework called CyberStrikeAI distributed tasks among AI agents. The other two involved human-directed attacks with AI assisting planning or tool development.[1], [2]
Target lists do not establish victim totals
The first environment used 44 agents with different tasks. Researchers identified at least 19 organizations in target-validation activity, including at least 16 in Japan. The environment contained reports in a common format and associated evidence indexes. The total number of successfully compromised organizations remains uncertain. The investigation distinguishes traces of target validation from systems known to have been breached. Japanese-language fake login pages were also among the recovered material.[1]
Human direction remained different across cases
In the second environment, a human operator used language-model assistance over 24 days spanning August and September. In the third, help with coding and troubleshooting was linked to at least six campaigns between January and October. Researchers did not establish that one attacker controlled all three environments. The underlying models were not conclusively identified either. The findings therefore describe different levels of assistance rather than the same degree of autonomy across every attack.[1]
