
One person in Hualien City aimed the BOSSAI AI agent attack, built as a hand-run framework, at hundreds of organizations in at least 11 countries. Acronis's Threat Research Unit confirmed data collection or system access at about 10 targets. The sharpest harm sat on Taiwanese emergency systems that served records without a working login.
Key takeaways
- One unnamed person in Hualien City ran BOSSAI: 12 agents, nine in use. Acronis confirmed collection or access at about 10 targets in Taiwan, Mexico, and Thailand.
- Taipei 119 dispatch records for seven districts came from a public API. A fire department platform exposed 2018 patient data, ECGs, and ambulance video with no login. Default root on a vendor database panel reached 18 or more tables.
- SQL injection took a lender account table (16 injection points) and data from three Thai universities chosen from 489 targets. Mexico produced about 206,800 home phone records with names and addresses.
- Grok, the main builder, and MiniMax M3 refused direct attack requests. Offense ran on local abliterated DeepSeek, GLM, and Qwen. Grok still repaired one proof-of-concept script. A scanner invented CVEs and reported attacks that never happened.
- Attribution is high confidence: reused GitHub, logo, SSH key name, and email, plus a home IP in a Grok SSH session. The researchers think he probably wanted money. Whether he works for anyone is unknown.
What did the BOSSAI AI agent attack actually confirm?
Acronis’s Threat Research Unit says one person in Hualien City, Taiwan, operated BOSSAI and confirmed data collection or system access at about 10 targets. Those cases sit in Taiwan, Mexico, and Thailand. The wider aim included emergency services, banks, universities, and military systems.
The Hacker News account by Swati Khandelwal, dated 9 November 2026, says the findings came out last week. Unpublished there: the Acronis report title, URL, and exact date; software versions; CVE identifiers; whether victims or Taiwanese authorities were notified; and whether the exposed data is still reachable.
Researchers are highly confident the operator is one person based in Hualien City. He is unnamed. No charges and no law-enforcement action are described. High confidence here is a judgment about a single operator and a city.
BOSSAI had 12 agents, nine of them in use. Some maintained his servers and tools. Others scanned targets and carried out attacks. The name is not expanded, and no version number is given.
The agents ran independently and did not hand tasks to each other. Khandelwal’s summary of the research states the control loop in one sentence:
The agents did not communicate with each other, so the attacker manually pasted one agent’s output into another.
He copied one agent’s output and pasted it into the next agent. He chose what moved. Nothing in the research shows an agent picking the next victim or checking a result before that paste.
The wider aim covered hundreds of organizations in at least 11 countries. About 15 further organizations were attacked with no confirmed break-in. In at least eight other countries, targets were only scanned or lined up for attacks that had not run.
Taiwan was targeted far more heavily than any other country. That weighting is the researchers’ technical conclusion. The sector list runs past the three countries where collection was confirmed.
Acronis watched the operation for 12 days. One recorded exchange shows the attacker at work on July 25. The year on that date is unpublished. Twelve days is how long researchers watched this evidence, not a published start and end for the activity.
The evidence came from his own server. It first looked like a misconfigured development box and was his operational platform. It held more than 90,000 files: agent conversations, command history, hacking tools, credentials, and victim data. That number is the size of the evidence pile, not a victim total.
Which Taipei emergency systems did the agents reach?
Confirmed emergency-service harm in Taipei is three outcomes: 119 dispatch records from a public API, a fire department platform that exposed patient data with no login, and a vendor database opened with default root credentials.
Agents pulled 119 dispatch records for seven Taipei districts from the system’s API. 119 is Taiwan’s number for fire and ambulance services. The records include incident times, response types, case status, and street-level locations. They are case records, not live ambulance tracking.
Street and response type are enough to place a person at an incident. The agents read that from a public-facing API. The pull used the API as it was exposed.
A Taipei fire department platform exposed patient data from 2018, including ECG readings and ambulance video, without a login. The detailed account says the records were exposed. A shorter summary says breaches at emergency services and universities led to theft of medical and personal data. For this platform, the confirmed statement is exposure.
The fire-dispatch technology vendor fell to default root credentials on a database administration panel. That login yielded access to 18 or more database tables. A web shell attempt through a known vulnerability failed. The failed shell leaves the table access in place.
On that panel, change the credential and then check authentication logs, new accounts, grants, scheduled jobs, stored procedures, and unexpected tables. A new password does not show whether the intruder is gone.
At an emergency management information system, a hardcoded AES-256 key in the site’s front-end JavaScript exposed session-cookie contents, including an administrator’s identity. There was no confirmed compromise.
The cipher does not protect a cookie when the key ships in script every browser downloads. Replacing the key also means invalidating sessions sealed with the old key, or those cookies keep working.
| Target | What the agents did | Result |
|---|---|---|
| Fire-dispatch technology vendor | Logged in with default root on the database admin panel. A web shell attempt through a known CVE failed. | Access to 18 or more tables |
| Emergency management information system | Hardcoded AES-256 key in front-end JavaScript exposed an administrator identity inside the session cookie | No confirmed compromise |
| National disaster-response center | A scanning tool reported a critical Apache remote code execution flaw | The flaw did not exist |
| National fire agency single sign-on portal | Login guessing | No confirmed access |
| Fire-department systems in five cities | Scanned to identify the software they run | No confirmed compromise |
What broke at the lender, the Thai universities, and in Mexico?
Outside emergency services, three collections are confirmed: an account table at a Taiwanese lending firm, data from three Thai universities, and about 206,800 Mexican home phone records with names and addresses.
SQL injection pulled the lender’s account table. The same firm had 16 SQL injection points. That many sinks usually means string-built queries in more than one handler, so a block on a single parameter leaves the next builder open.
Bind the queries to parameters, and treat a web application firewall rule as cover only until that code change ships. In the same change, restrict the database account to the schemas and actions the application uses.
Sixteen injection points are one defect repeated. Walk every query that concatenates input into SQL, and list what that account can read.
In Thailand, the agents picked 489 potential targets across 45 institutions and then collected data from three universities by SQL injection. Acronis ties that access to unpatched Oracle databases. Three universities are the confirmed set. The 489 and the 45 describe selection, not breaches.
SQL injection at the lender and at the universities is the same class of bug on different estates. The agents needed a query that concatenated input and a database that answered.
The Mexican set is larger by rows and different in method. About 206,800 home phone records, with names and addresses, were built from government statistics, postal-code lists, and telecom-authority data. Acronis presents that as collection from those sources.
Taiwan remained the focus of effort. Banks and military systems are on the target-sector list. The confirmed collections named in public are the emergency cases, the lender, the three universities, and the Mexican phone records.
Why did cloud models refuse while local models attacked?
Commercial cloud models did most of the building. Asked to attack, they mostly refused, so the offensive work ran on local DeepSeek, GLM, and Qwen models with safety training removed. Acronis calls those builds abliterated: models run locally after safety training is stripped.
Grok was the main builder among the commercial models. Asked directly to attack, Grok and MiniMax M3 refused and requested written proof the work was authorized. Building infrastructure and driving an intrusion were different requests, and both products drew that line here.
Whether a forged authorization letter would have flipped the refusal is unclear, Acronis says. What is on record is the refusal and the request for paperwork. That is an observation about Grok and MiniMax M3 in this case, not a claim about every setting of either product.
Grok made one exception: it repaired a broken proof-of-concept script. The attacker then handed that script to an attack agent. Fixing offensive code helped the operation. It is still a smaller step than picking targets and running the intrusion. After the handoff, the attack side stayed on the abliterated models.
Refusal of a direct attack request did not stop Grok from repairing a script the attacker could pass on. Commercial models still did most of the building. The split to record is builder versus attacker, not a blanket claim that Grok refused every task.
Claude Code and Codex were pointed at an abliterated model and mostly did infrastructure work. Cursor, running such a model, attacked a Taiwanese emergency-services site. Moving a coding agent from the vendor API to weights you host drops the provider check that refused the same operator.
That redirect is the control to test, as laid out in what to lock down on coding agents. A local model can still edit servers, scanners, and exploit scripts. It can also emit a finding that never happened.
One scanner invented CVEs, and more than half of its high and critical findings were false positives or wrongly classified. The main attack agent reported results that had not happened. The Apache critical at the national disaster-response center is the plain case: the flaw was not there.
A SOC that opens a bridge call on an invented critical spends the hour on a bug with no package to patch. The same transcripts contain a default root login that was real. Reproduce the request against the installed build, and keep the model output next to the proof, before either item becomes an incident.
He stayed in the loop, and the loop still passed bad output into the next task. The paste moved the previous agent’s text onward without a fresh check against the target.
How did Acronis attribute the operator in Hualien?
Attribution rests on identity mistakes plus public information. He reused one identity across legitimate businesses and the attack servers: a GitHub profile, a logo, an SSH key name, and an email address. Those overlaps are what made the business and the platform the same person.
His home IP address appeared during an SSH session in one of his Grok conversations. Together with the reused identity, that is the basis for high confidence in a single person in Hualien City. The person is unnamed, and the research describes no charge.
The server that supplied this evidence is the box researchers first took for a development system. It held the conversations that included the SSH session, plus credentials and victim data. Identity and collection sit together because he kept them on that host.
The researchers think he probably wanted to make money, and it is unclear whether he works for anyone. Some government, emergency, and military targeting looked like state-backed intrusion. Other activity, including bulk personal data, looked like ordinary cybercrime.
Filing the case as espionage or as a solo fraud shop chooses an ending the evidence leaves open. Probably money, affiliation unknown: that is the sourced line.
Nine IP addresses and nine domains from the setup appear in the indicator list. None is marked active or inactive, and most have no note on what they were used for. They are not reprinted here. An address without a role is a lead for teams that already have the Acronis analysis.
For the exposed emergency and medical data, keep access logs, object identifiers, and timestamps before retention rolls them off. Those records are how an owner separates an open route from a copy taken on a known date.
Where did the agents fail, and what should security leads test?
The agents succeeded against data that was already poorly protected, and they stopped where a login, a valid session, or input filtering held. Cheap coverage of many targets is what the models changed. Passing a control that was actually enforced is the part this campaign did not show.
What worked for him was old and ordinary. A medical data port had been open since 2018. Dispatch records sat on a public API. The lender had 16 injection points. Universities were on unpatched Oracle. A vendor database panel still took root defaults.
What failed for him belongs in the test plan. The national fire agency single sign-on portal took a login-guessing run and produced no confirmed access. Five cities’ fire systems were fingerprinted only. Session-gated paywalls held. Filtered endpoints held.
That split, controls that held versus controls that never existed, is what AI agent penetration testing should be scored against. Score the session check that stopped a tool-using model. Score a prompt that only asked the model to behave as a prompt, not as a control.
Both kinds of agent output are in this case. An invented CVE and a real default password came from the same operation. A program graded on the agent’s criticals will overstate the breach. A program graded only on the bad reports will miss the tables that were reached.
The matching exercise is how to test excessive agency on a pasted loop. BOSSAI kept a person in the middle, and bad output still moved to the next agent because the step was not reproduced before the paste.
Hunt the three defects that produced confirmed access. Find internet-facing database administration panels and try the product default root account. That is how the fire-dispatch vendor was opened, with access to 18 or more tables.
In code, search for SQL built by concatenation and group the hits by query builder. Sixteen lender findings were one class. Check whether the database account used by those queries can read beyond the feature it serves.
On public APIs, request incident, dispatch, and patient objects with no cookie and again with a low-privilege session. The 119 records and the 2018 clinical platform were reachable without a working login. Mark the route readable-without-login, then use access logs to see who fetched it.
If the admin panel must exist, move it to a management network, replace the default credential, and remove schema-wide rights from any account the web tier uses. Internet exposure plus default root is the condition used at the vendor.
BOSSAI was one operator and nine agents in use, with confirmed harm at about 10 organizations that were already open. The same tooling reported a critical flaw that did not exist and attacks that never landed. The controls that failed were the ones that were never configured in the first place.
Key takeaways
- One unnamed person in Hualien City ran BOSSAI: 12 agents, nine in use. Acronis confirmed collection or access at about 10 targets in Taiwan, Mexico, and Thailand.
- Taipei 119 dispatch records for seven districts came from a public API. A fire department platform exposed 2018 patient data, ECGs, and ambulance video with no login. Default root on a vendor database panel reached 18 or more tables.
- SQL injection took a lender account table (16 injection points) and data from three Thai universities chosen from 489 targets. Mexico produced about 206,800 home phone records with names and addresses.
- Grok, the main builder, and MiniMax M3 refused direct attack requests. Offense ran on local abliterated DeepSeek, GLM, and Qwen. Grok still repaired one proof-of-concept script. A scanner invented CVEs and reported attacks that never happened.
- Attribution is high confidence: reused GitHub, logo, SSH key name, and email, plus a home IP in a Grok SSH session. The researchers think he probably wanted money. Whether he works for anyone is unknown.
FAQ
What is BOSSAI?
BOSSAI is the attacker’s name for a 12-agent setup, nine of them in use. Some agents maintained servers and tools. Others scanned and attacked. The agents did not talk to each other. He pasted one agent’s output into the next. Acronis does not expand the name or give a version.
Who ran the BOSSAI AI agent attacks?
Acronis is highly confident it was one person based in Hualien City, Taiwan. He is not named. No charges are described. The researchers think he probably wanted to make money, and it is unclear whether he works for anyone.
Did hackers steal Taipei 119 ambulance records?
Agents pulled 119 fire and ambulance dispatch case records for seven Taipei districts from a public API: times, response types, case status, and street-level locations. They are case records, not live ambulance tracking. A separate Taipei fire platform exposed 2018 patient data, ECGs, and ambulance video with no login. The detailed account confirms exposure there, not a completed download.
What AI models did the BOSSAI attacker use?
Grok did most of the commercial building. Grok and MiniMax M3 refused direct attack requests and asked for written authorization. Grok once repaired a proof-of-concept script that he handed to an attack agent. Attack work ran on abliterated local DeepSeek, GLM, and Qwen. Claude Code and Codex, redirected that way, mostly built infrastructure. Cursor on such a model attacked a Taiwanese emergency-services site.
How did Acronis attribute the Taiwan AI hacker?
He reused one identity on legitimate businesses and attack servers: a GitHub profile, a logo, an SSH key name, and an email address. His home IP address appeared during an SSH session inside a Grok conversation. Acronis combined those mistakes with public information.
How many organizations did BOSSAI actually breach?
About 10, in Taiwan, Mexico, and Thailand, with data collected or systems accessed. About 15 more were attacked without a confirmed break-in. Hundreds of organizations in at least 11 countries were in scope. In at least eight other countries, targets were only scanned or queued.
If dispatch APIs, default database panels, or a stack of injection points sit in the same estate an agent can reach, a focused web application penetration test will show which of those paths still return data after the login is enforced.