Acronis Details BOSSAI AI-Agent Campaign Targeting 11 Countries

Summary
Acronis TRU describes BOSSAI, a 12-agent AI attack mesh operated by one person and used against organizations in 11 countries. The investigation reports confirmed data access or exfiltration at roughly 10 targets, including emergency services and universities.
Key points
- Acronis TRU says it analyzed more than 90,000 files and 9GB of evidence, including orchestration logs, agent conversations, commands, tools and victim data.
- BOSSAI comprised 12 agents, nine active; the operator manually relayed information between agents rather than enabling agent-to-agent communication.
- Offensive agents primarily used locally run models with safety alignment removed, while commercial models repeatedly refused explicit attack requests.
- The operator combined AI agents and MCP servers with tools including nmap, sqlmap, Metasploit and Ghidra; some agents ran autonomously with actions auto-approved.
- The investigation reports confirmed system access or data exfiltration at roughly 10 targets across three countries, while other attacks and reconnaissance remained unconfirmed.
- In Taiwan, exposed emergency-services systems provided access to patient-care records and dispatch case data; the article also reports extracted data from a Taiwanese lending firm.
- In Thailand, the investigation found confirmed SQL injection and data extraction at three universities. Separately, more than half of HalberdStrike's high-severity and critical findings were false positives or misclassifications.
Article Details
- Attack Vectors
- Automated vulnerability scanning and exploitation of public-facing web applications, including confirmed SQL injection against Taiwanese and Thai targets.
- Access to emergency-services medical records and dispatch data exposed through unauthenticated platforms and APIs.
- Brute-force attempts against a fire-agency single sign-on portal and credential-stuffing attempts against a Mexican site's WordPress XML-RPC endpoint; neither had confirmed successful logins.
- A Taiwanese fire-dispatch technology vendor's database administration panel was accessed using default root credentials.
- The operator used AI agents and pentesting tools to enumerate targets, analyze scan results, write scripts, and carry out exploitation and data collection.
- Defensive Notes
- Require authentication for medical records, emergency-dispatch data, and other sensitive API resources; the article reports these were accessible without credentials.
- Remediate SQL injection in production applications and secure database systems, including legacy deployments.
- Enforce server-side payment-card validation; agents demonstrated that one provider accepted a card despite an incorrect CVV.
- Replace default credentials and protect login endpoints against brute-force and credential-stuffing attempts.
- Apply scope controls and human review to autonomous security agents; the article found a pentesting tool's scope enforcement disabled and its risk tiers set to auto-approve.
- Do not rely on obscurity as a defense; the article reports that automated scanning rapidly identified targets across many organizations.
Indicators of compromise
| Type | Indicator | Context |
|---|---|---|
| DOMAIN | 888[.]bossjy[.]kdns[.]fr | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| DOMAIN | attacker-collector[.]com | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| DOMAIN | boss[.]jytian[.]cc[.]cd | Listed in the article's BOSSAI IOC section; the article links the operator's Jy Technical Team branding to the attack infrastructure. |
| DOMAIN | c[.]xs1[.]ccwu[.]cc | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| DOMAIN | ssh[.]gputw[.]ai | Listed in the article's BOSSAI IOC section; the article describes an operator GPU cluster connected to the attack platform. |
| DOMAIN | vps[.]xs1[.]ccwu[.]cc | Listed as a BOSSAI IOC; the article also describes an operator instruction to configure a VPS address with a closely matching hostname. |
| DOMAIN | xss[.]xs1[.]ccwu[.]cc | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 112[.]121[.]185[.]218 | Operator-controlled vps12 attack platform hosting OpenClaw C2 and offensive tooling. |
| IPV4 | 156[.]240[.]109[.]26 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 202[.]160[.]76[.]153 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 202[.]160[.]76[.]154 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 202[.]160[.]76[.]155 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 202[.]160[.]76[.]156 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 202[.]160[.]76[.]157 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 202[.]160[.]76[.]174 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
| IPV4 | 49[.]158[.]237[.]136 | Listed in the article's BOSSAI IOC section as campaign-related infrastructure; its specific role is not described. |
MITRE ATT&CK
T1046 · Network Service DiscoveryThe attack agents ran nmap port scans and enumerated target services.T1059.004 · Unix ShellThe operator used bash scripts and terminal commands for target scanning, exploitation, and data-collection workflows.T1078 · Valid AccountsAgents logged into a database administration panel using default root credentials.T1090 · ProxyAttack traffic was routed through an operator-controlled SOCKS5 pool and commercial datacenter proxies.T1110.001 · Password GuessingThe operator attempted brute-force access against a Taiwanese fire-agency single sign-on portal and a consumer finance authentication system.T1110.004 · Credential StuffingThe operator attempted credential stuffing against a Mexican site's WordPress XML-RPC endpoint using more than 2,600 username-password combinations.T1119 · Automated CollectionAgents automated collection, including scraping directories and extracting records from databases and APIs.T1190 · Exploit Public-Facing ApplicationThe operator's agents exploited SQL injection and other weaknesses in public-facing applications.T1595.002 · Vulnerability ScanningThe operator used automated vulnerability scanning against target organizations, including web applications and university systems.
People
Threat Actors
Vendors
Anthropiccampaigns, using all the necessary tools, and with the same speed of attack demonstrated across Anthropic's reports of multiple threat clusters operating globally.MiniMaxcomplying only once to fix a broken POC script, while OpenClaw (17 sessions, running commercial MiniMax M3) refused every offensive request, citing Computer Crime Law in response. Despite this experimentation,OpenAIconfigured the OpenClaw-to-LINE connection, as well as a few other tasks as needed. Claude Code and OpenAI's Codex CLI were notable exceptions to the commercial-for-infrastructure pattern: The operator ran shimsxAI- this seems incomplete."Of the several agents participating in infrastructure building, Grok, running on xAI's own cloud service via Grok 4.5, was the primary builder. Across 14 sessions, Grok assembled most of the
Products
Claude Code12 AI agents, each with different roles, capabilities and usage frequency. Builders (Grok, OpenClaw, Claude Code, Codex) assembled and maintained the mesh: MCP wiring, DNS, proxy configuration, messaging-appCodex CLIthe OpenClaw-to-LINE connection, as well as a few other tasks as needed. Claude Code and OpenAI's Codex CLI were notable exceptions to the commercial-for-infrastructure pattern: The operator ran shims thatCursorthe mesh: MCP wiring, DNS, proxy configuration, messaging-app gateways. Attackers (Hermes above all, plus Cursor, Trae and CyberStrike) scanned targets and built exploits. A dedicated scanning layer (the agenticDeepSeek-V4tested on real tasks, and either kept or replaced within hours. The final default was an abliterated DeepSeek-V4 variant the operator called DS4, running on locally hosted GPUs with 64K context.GLMexecuted attacks, ran almost exclusively on abliterated models: stripped-down variants of DeepSeek, GLM, and Qwen with safety alignment removed.Grok 4.5several agents participating in infrastructure building, Grok, running on xAI's own cloud service via Grok 4.5, was the primary builder. Across 14 sessions, Grok assembled most of the mesh's connective tissue. ItMiniMax M3complying only once to fix a broken POC script, while OpenClaw (17 sessions, running commercial MiniMax M3) refused every offensive request, citing Computer Crime Law in response. Despite this experimentation,OpenClawTraditional Chinese, primarily typed directly into each agent's interface. Two of the agents, Hermes and OpenClaw, had been set up with an integration to messaging apps. A WeChat gateway routed to Hermes, the primaryQwen 3.5 122Bmodels frequently. Early configurations included a rented Gemma model served from a cloud GPU and Qwen 3.5 122B on the operator's own hardware, though the exact deployment order between these early models isTraeMCP wiring, DNS, proxy configuration, messaging-app gateways. Attackers (Hermes above all, plus Cursor, Trae and CyberStrike) scanned targets and built exploits. A dedicated scanning layer (the agentic HalberdStrike
Tools
Acunetix Web Vulnerability Scannerand built exploits. A dedicated scanning layer (the agentic HalberdStrike and a cracked Acunetix Web Vulnerability Scanner installation) fed structured vulnerability reports into the attack pipeline.CyberStrikeDNS, proxy configuration, messaging-app gateways. Attackers (Hermes above all, plus Cursor, Trae and CyberStrike) scanned targets and built exploits. A dedicated scanning layer (the agentic HalberdStrike and aFOFAThe heaviest individual tool invocations were sqlmap (appearing in 561 messages across sessions), FOFA queries for asset discovery (432), nmap port scans (271), Shodan lookups (213), and nikto web-serverHalberdStrikeTrae and CyberStrike) scanned targets and built exploits. A dedicated scanning layer (the agentic HalberdStrike and a cracked Acunetix Web Vulnerability Scanner installation) fed structured vulnerability reportsHexStrike AIBy the time we'd investigated the MCP stack, this had been trimmed to three: HexStrike AI (a custom cybersecurity automation server), a Kali MCP server and a Ghidra bridge. These tools are then invoked autonomously byMetasploit Frameworkbrowser automation via Playwright, and an MCP-wrapped Kali Linux instance exposing nmap, sqlmap, Metasploit, Hydra, and more.Nmapvia Ghidra and Radare2, browser automation via Playwright, and an MCP-wrapped Kali Linux instance exposing nmap, sqlmap, Metasploit, Hydra, and more.Oktos Adaptive OffensiveZen Terminal on port 18080, Kali MCP server on port 5001, HexStrike AI on port 8000, and the Oktos Adaptive Offensive post-exploitation framework on port 8443. Port 3000 served "Jyæè¡åé" ("Jy TechnicalShodanoperator asked Hermes to find a "website that concentrates Mexican residential phone data." Hermes ran a Shodan search for APIs and stumbled upon a Honduran military host, which it then briefly attempted to exploitsqlmapand Radare2, browser automation via Playwright, and an MCP-wrapped Kali Linux instance exposing nmap, sqlmap, Metasploit, Hydra, and more.
Countries
Australialevel. Infrastructure-search queries targeted universities in the United States, the United Kingdom, Australia, Sri Lanka, and Singapore. Corporate and government-adjacent systems in China and Vietnam were probedCanadaURLs across Hong Kong (financial firms and private clubs), Indonesia (retail and sportswear sites), and Canada (Chinese-diaspora e-commerce and media outlets) sat in a project queue that recorded zero findings andChinaVesta with SMTP, IMAP and POP3 services. Behind all of this was the cn-4090 GPU cluster in Guangzhou, China, running uncensored models across 192GB of VRAM, tunneled to vps12 via autossh on eight ports.HondurasFigure 8: Hermes notes the Honduras reconnaissance in one of its skill files, as a lesson learned. Stating that “it contains phone numbers for Honduran military personnel, not Mexican White Pages,” and concluding itHong Kongautonomous attack platform but never scanned at the time of investigation. Fifty-seven URLs across Hong Kong (financial firms and private clubs), Indonesia (retail and sportswear sites), and CanadaIndonesiaat the time of investigation. Fifty-seven URLs across Hong Kong (financial firms and private clubs), Indonesia (retail and sportswear sites), and Canada (Chinese-diaspora e-commerce and media outlets) sat in aMexicoThe Mexico operation had a different character: mass collection of residential personally identifiable information. The operator tasked agents with building a phone directory covering three cities and 12 affluentSingaporequeries targeted universities in the United States, the United Kingdom, Australia, Sri Lanka, and Singapore. Corporate and government-adjacent systems in China and Vietnam were probed through bulk scanningSri Lankaqueries targeted universities in the United States, the United Kingdom, Australia, Sri Lanka, and Singapore. Corporate and government-adjacent systems in China and Vietnam were probed through bulkTaiwanBOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency servicesThailandgatekeeper," Minimax M3 replied, and later added "attacking without written authorization = violating Thailand Computer Crime Act Sections 5-7 (up to 5 years). You need to protect yourself too -- withoutUnited Kingdomreconnaissance level. Infrastructure-search queries targeted universities in the United States, the United Kingdom, Australia, Sri Lanka, and Singapore. Corporate and government-adjacent systems in China and VietnamUnited Stateswere touched at the reconnaissance level. Infrastructure-search queries targeted universities in the United States, the United Kingdom, Australia, Sri Lanka, and Singapore. Corporate and government-adjacent systems inVietnamKingdom, Australia, Sri Lanka, and Singapore. Corporate and government-adjacent systems in China and Vietnam were probed through bulk scanning campaigns. Japanese targets had search queries prepared but not yet
Industries
E-commercegovernment emergency services, banks and lenders, universities, military systems, phone directories, e-commerce platforms, and gambling sites in eleven countries. It’s unclear if the attacker has any specificEmergency ServicesBOSSAI: Inside the AI attack mesh that targeted Taiwan’s emergency servicesGamblingbanks and lenders, universities, military systems, phone directories, e-commerce platforms, and gambling sites in eleven countries. It’s unclear if the attacker has any specific affiliation or goal beyondGovernmentUsing the BOSSAI mesh, the operator targeted hundreds of organizations spanning government emergency services, banks and lenders, universities, military systems, phone directories, e-commerce platforms, and gamblingMilitaryhundreds of organizations spanning government emergency services, banks and lenders, universities, military systems, phone directories, e-commerce platforms, and gambling sites in eleven countries. It’s unclear
