Acronis Details BossAI Automated Attack Framework Targeting Taiwan Emergency Services and Global Entities

The Acronis Threat Research Unit disclosed an automated attack framework named BossAI, which targeted Taiwan's emergency services systems and hundreds of organizations across 11 countries, allegedly operated by an individual in Hualien City, Taiwan. Researchers retrieved approximately 9GB of data across 90,000 files from a misconfigured development environment to reconstruct the attack workflow. Although designed around 12 AI agents, only nine were active during the investigation, with no direct inter-agent communication, requiring manual transfer of outputs. The operator used custom Skills to preserve exploit procedures and disabled safety boundaries on local models. However, over half of the flagged critical vulnerabilities were false positives, showing that automated attacks remain constrained by model reliability, and targets with robust authentication, session controls, and network filtering successfully prevented compromise.

Key takeaways

• Acronis found BossAI was operated by an actor believed based in Hualien, targeting Taiwan emergency systems and hundreds of global entities via nine active AI agents and MCP.
• The operator utilized custom Skills to store exploit flows, with a Skill targeting Taiwan's fire dispatch and emergency management APIs used most frequently alongside HalberdStrike.
• Safety boundaries and tool scopes were removed to automate operations, though over half of reported high-severity vulnerabilities were false positives or misclassifications.
• Acronis noted AI agents reduce the cost of large-scale vulnerability discovery, but systems with proper authentication, session controls, and network filtering successfully prevented compromise.