Get a free audit
Apple credit Apple credits the researcher who runs our AI penetration testing.

Field Notes / AI Threats

One person, twelve AI agents, Taipei’s ambulance records. BOSSAI used holes older than the AI.

Acronis has published the logs of BOSSAI, a lone operator in Taiwan who ran a mesh of AI agents against hundreds of organisations in eleven countries. He reached fire-department patient data with no login required, and a university database through a SQL injection in Oracle 9.2. Here is how the mesh worked, what it got wrong, and why the defence is the boring list you have been putting off.

Author
Red Team Partners
Read
9 MIN READ
Filed
10 Nov 2026
A GPU workstation of the kind BOSSAI used to run its uncensored attack models at home.

01 What Acronis found

cronis’s Threat Research Unit stumbled on what looked like a misconfigured development environment. Inside sat more than 90,000 files and 9GB of evidence: agent conversations, command histories, exploit tools and stolen victim data Acronis TRU 2026 . The attacker had left the workshop open to the internet.

Acronis attributes the campaign, with high confidence, to a lone operator in Hualien City, Taiwan GBHackers 2026 . It labels the campaign an "AI-augmented threat" and finds no clear political goal. The targets were a mix: government emergency services, banks and lenders, universities, military systems, phone directories, e-commerce and gambling sites.

BOSSAI in numbersFigure
AI agents in the mesh12 (9 active, 3 configured but unused)
Countries targeted11
Messages in the main attack agent18,311 across about 95 sessions
Terminal commands runOver 3,300
Models tried by that agentAt least 9 in 19 days
Files left exposed by the attackerOver 90,000 (9GB)

02 How the agent mesh worked

Acronis splits the agents into two teams Acronis TRU 2026 . Builders, including Grok, OpenClaw, Claude Code and Codex, assembled and maintained the setup. Attackers, led by an agent called Hermes alongside Cursor, Trae and CyberStrike, scanned targets and wrote exploits. An agentic scanner called HalberdStrike and a cracked copy of the Acunetix web scanner did the bulk scanning.

The agents never talked to each other. The operator sat in the middle, pasting one agent’s output into the next, and gave orders in Traditional Chinese over WeChat and LINE. At its peak the mesh had 13 tool servers attached, wiring the agents straight into nmap, sqlmap, Metasploit and the Ghidra reverse-engineering suite.

Hermes was given a standing instruction, translated by Acronis: take full decision-making authority, do not ask, report back only when a vulnerability is confirmed exploitable. Auto-approval flags were switched on, so no command waited for a human. In practice that meant 561 sqlmap messages, 432 FOFA searches, 271 nmap runs and 213 Shodan queries.

03 What the agents reached

Look at the confirmed hits and you see a list any penetration tester would recognise from a 2012 report.

  • Taipei fire-department patient care. ECG readings, ambulance video and patient vital signs, reachable without credentials, with records back to 2018 GBHackers 2026 .
  • Taipei 119 dispatch. Incident times, response types and street-level locations across seven districts, through a public API. Acronis stresses these were case records; live ambulance positions were not exposed.
  • A Taiwanese financial firm. Sixteen SQL-injection points in one application GBHackers 2026 .
  • A Thai university. Over 50 subdomains enumerated, an Oracle 9.2.0.4.0 server found, and a blind SQL injection that returned a 2,563-row administrator table before the agent began on student records Acronis TRU 2026 .

Oracle 9.2 dates from 2002. The fire-department platform needed no password. The dispatch data sat on an API anyone could call. An AI was not required for any of it. What the AI changed was reach: one person found these systems among hundreds of targets in eleven countries.

The agents also got plenty wrong. Over half of HalberdStrike’s high and critical findings were false positives or misclassified, and some models invented CVEs. One reported a failed connection as a critical remote code execution flaw Acronis TRU 2026 . Hermes found a Honduran military host through Shodan, tried it, and then dropped it as "out of scope". Volume made up for accuracy.

04 What changes for defenders

Security teams have long relied on a quiet assumption: the forgotten system is safe because nobody is looking at it. An attacker with limited hours spends them on valuable targets. BOSSAI ends that. Agents do not get bored, and they will probe a fire-department reporting tool from 2018 as patiently as a bank.

Three practical shifts follow.

  • Your long tail is now in scope. Partner APIs, old subdomains, the reporting tool a supplier built in 2018. Each now gets tested by someone, often within days of going live.
  • Noise rises before damage does. BOSSAI’s agents produced bursts of scanning and injection attempts through commercial proxies and datacentre ranges. Alerts on that pattern buy you time.
  • Old fixes are still the fix. Every confirmed access in the report closes with authentication, patching, input validation or a changed password. None of it needs an AI product.

05 Six checks for this week

Cyber Press lists Acronis’s advice as patching internet-facing applications, fixing SQL injection, changing default credentials and closing unused ports Cyber Press 2026 . Here is that advice as a week of work for an IT lead. If you want an outside view first, start with a free security audit of what you expose.

This week
  • List every subdomain and public IP your organisation owns, including ones run by suppliers, and mark which ones nobody can name an owner for
  • Call every public API without a token and confirm it refuses. Partner and reporting APIs first, because those are the ones built without login
  • Find every database version facing the internet, directly or behind an app. Anything out of vendor support gets a patch date or a retirement date
  • Run an authenticated SQL-injection test against your oldest internet-facing application, the one built before your current team arrived
  • Change every default administrator credential on database panels, cameras, routers and admin consoles, and close ports nobody can justify
  • Set an alert for bursts of failed logins, injection payloads or scanner signatures from datacentre and proxy ranges, and route it to someone who reads it the same day

BOSSAI was caught because the attacker was careless with their own server. The next one will be more careful. The systems BOSSAI reached would still have been open. We run an external penetration test on your estate the way an agent mesh does, then do what the agents could not: prove which findings are real. In 14 days you get every exposed system we can reach, what an attacker gets from each, and the change that closes it. Start with your Threat Map, from $5,000. Your red team. Within reach.

Questions readers ask

What is BOSSAI?

BOSSAI is the name Acronis Threat Research Unit gave to an attack campaign it published on 4 November 2026. A single operator in Taiwan ran a mesh of 12 AI agents, nine of them active, against hundreds of organisations in eleven countries, including emergency services, lenders, universities, phone directories, e-commerce and gambling sites. Acronis calls it an AI-augmented threat. It found the operation through what first looked like a misconfigured development environment holding more than 90,000 files and 9GB of evidence.

Which AI models did the BOSSAI attacker use?

Acronis reports that commercial models, including Grok and MiniMax, refused offensive requests and asked the operator for proof that the work was authorised. The attack agents instead ran on "abliterated" versions of DeepSeek, GLM and Qwen, open models with their safety alignment stripped out, hosted on the operator’s own RTX 4090 GPUs. The main attack agent, Hermes, cycled through at least nine models in 19 days. Commercial coding tools such as Claude Code and Codex were used to build and maintain the setup.

What data did BOSSAI access in Taiwan?

According to the reporting on Acronis’s findings, a Taipei fire-department patient-care platform exposed ECG readings, ambulance video and patient vital signs without any credentials, with records going back to 2018. A Taipei 119 dispatch interface exposed incident times, response types and street-level locations across seven districts. Acronis notes these were case records; live ambulance positions were not exposed.

Are AI agents making cyberattacks more dangerous?

In BOSSAI they made attacks faster and wider. The techniques stayed old. The agents ran over 3,300 terminal commands and hundreds of sqlmap, nmap and Shodan jobs, but every confirmed access came through old weaknesses: unauthenticated APIs, SQL injection and unpatched databases. More than half of the scanner agent’s high and critical findings were false positives. The practical change for defenders is that a single person can now test every forgotten system you own, so the time between exposure and discovery shrinks.

How do I protect my organisation from AI-driven attacks like BOSSAI?

Fix what the agents exploited. Put authentication on every public API, including ones meant only for partners. Test internet-facing applications for SQL injection and patch or retire old database versions. Change default credentials and close ports nobody uses. Then watch for the pattern BOSSAI produced: bursts of automated scanning and injection attempts from proxies and datacentre IP ranges. An external attack-surface test shows you what an agent mesh would find first.

References

Sources

  1. Acronis Threat Research Unit (Kimhy, Pontiroli, Singha). BOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency services. 4 November 2026. acronis.com
  2. GBHackers. AI-powered BOSSAI attacks. 5 November 2026. gbhackers.com
  3. Cyber Press. Hacker unleashes BOSSAI agents. 5 November 2026. cyberpress.org