55/100

Moderate evidence

Cite
More
EmailLinkedInX

BOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency services

Evidence type
Investigative reporting
Sources
1 sourceJournalism
Published
3 min read

Introduction In recent weeks, the Acronis TRU team has identified and investigated an agentic, AI-driven campaign orchestrated by a lone operator based out of Taiwan, targeting hundreds of organizations in eleven countries and across sever…

BOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency services
"Nuclear Forensics" by IAEA Imagebank, CC BY-SA 2.0.

49Gain side

How the two sides scoreProblem scores higher

56Problem side

The gain

In Acronis's investigation of the BOSSAI campaign, commercial models including Grok 4.5 and MiniMax M3 refused every attack request, forcing the operator onto weaker local models.

The problem

Acronis researchers found a lone operator used a mesh of 12 AI agents to target hundreds of organizations in eleven countries, with confirmed breaches at universities and emergency services that exfiltrated medical data and PII.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

Acronis's Threat Research Unit uncovered BOSSAI, an attack platform run by one operator based in Taiwan. It joined 12 AI agents (nine active) to scan, exploit and build infrastructure against hundreds of organizations in eleven countries, including emergency services, banks, universities and military systems. Breaches at universities and emergency services providers led to confirmed theft of medical data and personally identifiable information.

Researchers found the platform after what looked like a misconfigured development environment. It held more than 90,000 files totalling 9GB, including agent conversations, command histories, exploit tooling, credentials and victim data. Acronis says this gives an unusually detailed view of a multi-agent offensive setup and proposes a new category, the AI-augmented threat, for lone attackers who rely on agents for technical work.

The campaign also shows limits. Commercial models refused offensive work and lectured the operator on computer crime laws, so attackers ran abliterated local models that made mistakes, hallucinated and produced more data than the operator could handle. Acronis warns that open-source pentesting platforms and unmonitored local models running with high autonomy could lower the bar for future attackers.

Main points

  1. The mesh had 12 agents: builders (Grok, OpenClaw, Claude Code, Codex) and attackers (Hermes, Cursor, Trae, CyberStrike), plus a scanning layer built on HalberdStrike and a cracked Acunetix scanner.
  2. Hermes, the main attack agent, ran about 95 sessions and produced 18,311 messages, 83.5% of them tool calls and results; it cycled through at least nine models over 19 days.
  3. In one case Hermes enumerated over 50 subdomains of a Thai university, confirmed a SQL injection and found an administrator table of 2,563 rows without further operator input.
  4. Over half of HalberdStrike's high-severity and critical findings were false positives or misclassifications, and one model flagged a failed connection as a critical RCE.
  5. Hermes was given full decision-making authority and passed over a Honduran military host because it judged the host out of scope, which Acronis calls ironic.
  6. Commercial models Grok 4.5 and MiniMax M3 refused attack requests, with MiniMax citing Computer Crime Law and asking for written authorization.
  7. The operator copied output between agents by hand; no automated agent-to-agent communication existed.
  8. The mesh ran on a local RTX 4090 GPU host serving abliterated models through llama.cpp, with eight VPS nodes and proxy layers for traffic.

The full story

The operator issued terse commands in Traditional Chinese, sometimes just a victim address, and relayed results between agents. In one documented chain, HalberdStrike scanned a Taiwanese emergency services target, Hermes consolidated the findings into a report with attack-path analysis, and Cursor, running an abliterated model, executed the exploit after the operator asked for help exploiting the vulnerabilities.

The offensive agents ran almost entirely on abliterated variants of DeepSeek, GLM and Qwen, with extra jailbreak prompts and auto-approve settings. Infrastructure agents mostly used commercial cloud models with safety filters intact, and the operator's attempts to jailbreak Grok and MiniMax M3 failed. Claude Code and Codex CLI were routed through shims to the abliterated DS4 model for some infrastructure work.

Hermes used reusable skill files that accumulated reconnaissance data. The most-used, gov-api-penetration, targeted Taiwanese fire-dispatch and emergency-management APIs and was invoked 51 times with 57 patches. Tool use included over 3,300 terminal commands and 129 Kali MCP calls. Acronis notes the operator's goals are unclear and the diversity of targets blurs the line between state-style espionage and cybercrime.

What this doesn’t fix

Acronis says it is unclear whether the commercial models would have complied had the operator supplied a fake engagement letter, and whether the Grok jailbreak text was ever used or worked.

Sources

1 on file
  1. Acronisacronis.com
    Journalism2026-11-04

Talk it over

Three questions to start on, two more if you have time
Grok 4.5 and MiniMax M3 refused every attack request, yet the operator turned to abliterated local models and still breached emergency services. How would you weigh the value of those refusals against the breaches that followed?

Start with a question

BOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency serviceshttps://truvace.com/item/bossai-inside-the-ai-attack-mesh-that-targeted-taiwans-emergency-services

  1. Grok 4.5 and MiniMax M3 refused every attack request, yet the operator turned to abliterated local models and still breached emergency services. How would you weigh the value of those refusals against the breaches that followed?

    Answer in the thread
  2. Acronis says it is unclear whether the commercial models would have complied if the operator had supplied a fake engagement letter. What finding would settle how much those refusals protect anyone?

    Answer in the thread
  3. The operator copied output between agents by hand, and over half of HalberdStrike's high-severity and critical findings were false positives. What should happen next, and who should decide, about open-source pentesting platforms and unmonitored local models?

    Answer in the thread
If you have more time
  • The case came to light after what looked like a misconfigured development environment holding more than 90,000 files. What does that suggest about how many similar setups go unseen, and what would you want to find out?

    Answer in the thread
  • If you ran the systems of an emergency services provider like the Taiwanese one that was breached, which details in Acronis's findings would change what you check first?

    Answer in the thread

What readers say