Acronis says one operator used 12 AI agents across 11 countries
Acronis traced a 12-agent AI attack network to one Taiwan-based operator, confirming breaches at emergency services and universities across 11 countries and documenting false-positive-heavy output.

Why it matters
- 18 November 2026: Acronis reported one Taiwan-based operator using 12 AI agents against hundreds of organizations in 11 countries.
- Nine of the 12 agents were active during the period Acronis examined.
- Confirmed university extraction reached three institutions, including one administrator table with 2,563 rows.
- Mexico collection produced about 206,800 phone records and roughly 124,000 additional directory records.
- More than half of one tool's high- and critical-severity findings were false positives.
Acronis's 18 November 2026 research identified one Taiwan-based operator using a network of 12 AI agents—software programs assigned to individual tasks—to attack hundreds of organizations across 11 countries, with confirmed breaches involving emergency services and universities.
The Acronis Threat Research Unit said: "AI doesn't necessarily make an attacker more sophisticated, but it can give a single individual significantly greater reach and operational capacity."
That assessment matters because Acronis documented a shift in scale rather than technical sophistication: one person directed 12 agents, nine active during the period examined, across hundreds of targets in 11 countries. The distinction matters.
Acronis found no clear political or ideological motive and assessed financial gain as the most likely objective. The case therefore offers a concrete benchmark for cyber-risk planning: one operator and 12 agents reached hundreds of organizations in 11 countries.
What did the 12-agent system do?
Acronis called the system BOSSAI, a mesh combining active and idle AI agents under one operator.
- Acronis said nine of the 12 agents were active during the period it examined; the system also included idle agents.
- Acronis said agents handled infrastructure management, scanning, report generation and exploitation.
- Acronis listed government emergency services, banks, lenders, universities, military systems, eCommerce platforms, gambling sites and phone directories among the operator's targets.
- Acronis confirmed theft of medical information and personally identifiable information (PII, data tied to individuals), including emergency patient care records and dispatch data.
- Acronis said the agents did not communicate directly. The operator manually moved outputs between tools and often pasted one agent's report into another to continue an attack.
- Acronis said the setup combined open-source penetration-testing platforms (security-testing tools) with model context protocol (MCP) servers connected to tools including Nmap, sqlmap, Metasploit, Hydra, Ghidra and browser automation frameworks.
- Acronis found the operation used a fleet of virtual private servers (VPS), local graphics processing units (GPUs) and rented computing resources.
Acronis described a workflow that separated reconnaissance, reporting and exploitation across distinct tasks. The human operator manually connected those stages, so coordination did not depend on direct agent-to-agent messaging.
Acronis found that the operator remained the hub for agent-to-agent handoffs, manually relaying outputs and pasting one report into the next tool. The research also found that some agents produced more output than one person could manage efficiently, making human supervision both the integration point and a capacity limit.
Why did the operator rely on local models?
Acronis said the offensive agents largely used locally run models with safety restrictions removed, while commercial AI models handled more infrastructure and coding work.
In several cases, commercial models refused to assist with unauthorized attacks and asked the operator to provide proof of legal authority.
Those refusals pushed the operator toward altered local models. Acronis found those models less dependable, sometimes producing false findings or missing targets, which created extra work and mistakes.
Acronis found that the refusals did not halt the campaign, but the resulting reliance on local systems produced false findings and missed targets. Commercial models supplied some infrastructure and coding work, while unrestricted local models carried more of the offensive load.
What did the agents access in Taiwan?
Acronis said Taiwan received the heaviest targeting. The campaign focused on emergency services, financial groups, government portals and a fire-dispatch technology supplier.
Acronis found that one emergency patient care records platform exposed medical data without credentials. The exposed records included electrocardiogram (ECG) readings, ambulance video and patient vital signs.
Acronis said agents accessed a city emergency dispatch system and extracted case records. Those records contained incident times, response types, status fields and street-level locations across several Taipei districts.
Acronis also documented SQL injection (database-query manipulation) against a lending services firm's Microsoft SQL Server backend. The researchers confirmed 16 injection points and account-information extraction.
Acronis found that a mobile payments card-binding process accepted cards even when a customer entered the wrong card verification value (CVV).
These cases give the campaign public-service significance because Acronis documented exposure of clinical measurements, ambulance video and location-linked dispatch data. The intrusion reached operational emergency records, not only account credentials.
How extensive were the university breaches?
Thailand was another major focus, particularly its universities. Acronis said the operator used infrastructure search tools to identify hundreds of potential university targets, then prioritized targets judged most likely to be vulnerable.
Acronis confirmed SQL injection and data extraction at three Thai universities. One breach exposed an administrator table containing 2,563 rows, while others exposed student information.
Acronis's finding separates confirmed data theft from exposed attack surface. Five universities had confirmed injection points, but only three had confirmed data extraction.
What did the Mexico operation collect?
Acronis said the operator built a residential directory covering multiple Mexican cities and affluent neighborhoods, producing about 206,800 phone records with names, addresses and postal codes.
Acronis also reported a separate scraping effort that gathered roughly 124,000 additional records from a directory site.
Acronis's Mexico findings show the same campaign supporting bulk aggregation as well as targeted intrusion. Unlike the confirmed university extraction counts in Thailand, the Mexico report quantified both collection efforts by record count: about 206,800 and roughly 124,000.
Where did the AI system fail?
Acronis found that AI did not remove the operator's need for judgment. Some agents hallucinated vulnerabilities, misread failed tests as successful discoveries and produced more output than one person could manage efficiently.
Acronis said one penetration-testing tool generated false positives—reported flaws that were not real—in more than half of its high- and critical-severity findings.
Acronis also documented an agent finding a Honduran military host during a Mexico-related search, then dropping the target after judging it outside the operation's scope.
Acronis's evidence shows why target count cannot serve as a proxy for successful compromise. Hundreds of potential targets, hallucinated vulnerabilities, prevalent false positives and a dropped military host coexisted within the same operation.
What does the case establish for defenders?
Acronis said the campaign showed that AI may not make an attacker more technically advanced, but it can let one individual investigate and attack many more targets in parallel.
That conclusion defines the near-term risk. Acronis documented one person moving outputs among 12 agents, nine active, while commercial-model refusals, hallucinated vulnerabilities, false positives and scope errors limited reliability.
For defenders, the immediate problem is separating exploitable evidence from the noise generated at this scale. Acronis recorded more than half of one tool's high- and critical-severity findings as false positives while one operator reached hundreds of targets in parallel. The validation gap, rather than a proven leap in attacker sophistication, is the measurable risk the research leaves behind.
Source: itbrief.com.au
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
718 articles