Hacker's BOSSAI operation used 12 AI agents to breach Taiwan emergency dispatch data and Thai universities, Acronis said
Cybersecurity firm Acronis said a single Taiwan-based operator ran an operation called BOSSAI, which combined 12 AI agents to scan, analyse and attack organisations in 11 countries. Acronis said the agents extracted case records from a city's 119 emergency-dispatch system covering seven Taipei districts. It also reported data extraction at three Thai universities and at a Taiwanese lending-services company. During a 12-day observation period, researchers confirmed unauthorised access or data extraction at approximately 10 targets across three countries. Commercial AI models such as Grok and MiniMax M3 were seen refusing attack requests, so the operator relied on locally hosted, safety-stripped variants of DeepSeek, GLM and Qwen for much of the direct exploitation.
What the AI did
A single operator used 12 AI agents, nine of them active, to scan, analyse and attack organisations, with an agent called Hermes doing most of the attacking. For much of the direct exploitation the operator used locally hosted versions of DeepSeek, GLM and Qwen that had been modified to remove their safety limits. Acronis said the agents pulled case records from a Taipei 119 emergency-dispatch system and used SQL injection, a common way of breaking into databases, to take account data from a lending company. Acronis also said more than half of one agent tool's serious vulnerability findings were false or misclassified.
First reported November 4, 2026 · Added to the register November 11, 2026 · 1 source
- Confirmed compromised targets
- About 10 across three countries
- Deployer
- Single Taiwan-based operator
- Model
- DeepSeek, GLM, Qwen variants, Hermes
- Affected
- Taiwanese lending-services company, City 119 emergency-dispatch system, Taipei, Three Thai universities
- Country
- Taiwan, Mexico, Thailand
- When it happened
- Not stated in the sources
- First reported
- November 4, 2026
What this means for you
Could this affect you?
If you run public-facing emergency, university, government or financial systems. Operators of public-facing emergency-service, university, government and financial systems, especially in Taiwan and Thailand, were targeted, along with people whose records sit in poorly protected databases.
What to check
- Require authentication and authorization checks on public APIs.
- Remove default credentials and fix SQL-injection flaws in older systems.
- Monitor for automated scanning of your systems.
- Limit tool permissions on your own AI agents and require human approval for sensitive actions.
Areas of your AI programme this touches
Every fact and its source (6)
- False positive rateMore than half of high-severity findings“more than half of HalberdStrike's high-severity and critical vulnerability findings were false positives or misclassifications”[1]
- Discovered byAcronis Threat Research Unit“Cybersecurity firm Acronis uncovered the operation, known as BOSSAI”[1]
- AI agents in operation12 (nine active)“BOSSAI consisted of 12 AI agents, nine of which were active”[1]
- Countries reached11“The campaign reached organizations across 11 countries”[1]
- Mexican phone records collectedAbout 206,800“One operation produced approximately 206,800 phone records associated with names, addresses, and postal codes”[1]
- Commercial models that refusedGrok and MiniMax M3“Researchers observed Grok and MiniMax M3 refusing offensive requests”[1]
Sources
- One Hacker Used 12 AI Agents to Target Taiwan's Emergency Servicesesecurityplanet.com · November 4, 2026
How this record is classified. Severity S (4/5): real harm to one organisation, data exposed, or a containment failure. OECD level: incident, an event in which an AI system led to actual harm. Evidence: Reported, meaning press reporting only.
The facts in At a glance and in Every fact and its source are quoted from the sources listed. The summary, What the AI did and What this means for you are written from those sources and checked against them automatically before publication. Records reflect what has been disclosed, not everything that has happened. Incidents are found through news feeds and the GDELT Project.

