Acronis uncovers AI agent mesh behind Taiwan service breaches

Acronis Threat Research Unit (TRU) says it uncovered BOSSAI, a network of 12 AI agents operated by a single person. During 12 days of observation, the operation scanned hundreds of organisations in 11 countries and caused about 10 confirmed breaches. An exposed server at 112.121.185[.]218 contained more than 90,000 files and 9GB of logs. TRU says the operator manually passed outputs between agents, using one group to build and maintain infrastructure and another to scan targets and run attacks.

After commercial models including MiniMax M3 and Grok 4.5 refused explicit hacking requests, the operator used locally hosted, uncensored versions of DeepSeek, GLM and Qwen. TRU says the main agent, Hermes, ran an abliterated DeepSeek-V4 model across 95 sessions, issuing more than 3,300 terminal commands and 561 sqlmap calls. The agents were not consistently reliable: HalberdStrike reported non-existent CVEs in more than half of its critical findings, according to the researchers.

Affected systems included a Taipei fire-department platform exposing 2018 patient ECG readings, ambulance videos and vital signs, as well as 119 dispatch records from seven Taipei districts. The operation also extracted account tables from a Taiwanese lender, bypassed CVV checks in a mobile payment app, and used SQL injection to obtain student and administrator tables from three Thai universities. In Mexico, agents collected about 206,800 residential phone records and 124,000 business listings.

TRU attributes the activity with high confidence to one individual in Hualien City, Taiwan, and says the agents failed where authentication, patching and API filtering were in place.