Acronis exposes operational details of the AI automated attack system BOSSAI, integrating 12 AI agents, MCP and Skills

Security vendor Acronis Threat Research Unit (TRU) disclosed that an AI attack architecture named BossAI targeted Taiwan's emergency services systems and hundreds of organizations in 11 countries worldwide, operated by an attacker believed to be located in Hualien City, Taiwan, integrating 12 AI agents, chaining penetration testing tools through MCP, and using a local AI model with safety restrictions removed to carry out attacks. Researchers obtained more than 90,000 files and approximately 9GB of data from a suspected misconfigured development environment, and used this to reconstruct how the architecture operated. During the investigation, 9 agents were actually operational, with no direct communication among them, and the attacker manually passed work results between them; the attacker also used Skills to save attack workflows and experience, among which gov-api-penetration, targeting Taiwan government fire dispatch and emergency management APIs, was used 51 times and modified 57 times. Hermes generated 18,311 messages across approximately 95 work stages, of which 83.5% were tool calls and execution results; more than half of the high-risk and critical weaknesses reported by HalberdStrike were false positives or misclassifications.