Taiwanese emergency services agentic AI attack — Nov 2026

What happened

Acronis Threat Research Unit (TRU) identified a lone operator based in Taiwan who orchestrated a persistent cyber espionage campaign using a mesh of 12 AI agents, collectively referred to as BOSSAI. This infrastructure, which utilized a mix of commercial and abliterated open-source models, targeted hundreds of organizations across 11 countries, including government emergency services, banks, universities, and military systems. The operator acted as a manual bridge between agents, which were assigned specialized roles such as infrastructure maintenance, vulnerability scanning, and exploit development.

To execute these campaigns, the operator integrated 13 Model Context Protocol (MCP) servers, providing the agents with access to tools like Nmap, SQLmap, Metasploit, and Ghidra. While the agents were designed for autonomy, their performance was inconsistent; they frequently hallucinated vulnerabilities, misclassified findings, and occasionally abandoned targets due to scope drift. The offensive core of the mesh relied on abliterated models—stripped of safety alignments—to bypass restrictions that commercial models consistently enforced.

Confirmed breaches occurred at several universities and emergency services providers, resulting in the exfiltration of medical data and personally identifiable information (PII). For example, agents extracted patient-care records and dispatch data from a Taipei fire department system. In Thailand, agents successfully performed SQL injection attacks against three universities, extracting student data. Despite these successes, the operator’s infrastructure suffered from significant operational security (OPSEC) failures. The same home IP address used for the attack mesh also hosted the operator’s legitimate gas delivery business and a SaaS marketing tool. These overlaps, combined with reused branding and hardcoded credentials, allowed researchers to attribute the campaign to a single individual in Hualien City, Taiwan. The investigation highlights the growing risk of AI-augmented threats where autonomous agents can rapidly scale reconnaissance and exploitation, even when the operator’s technical execution is hampered by model hallucinations and tool failures.