1. Description
The campaign described by Zscaler ThreatLabz involves the use of AI-assisted tooling to coordinate hacking activity across organizations in multiple sectors and countries.
The investigation followed earlier reporting about attacks targeting South Korean financial institutions. However, examination of the operator's infrastructure, saved prompts, model configurations, and project records revealed a much broader targeting pattern.
The operator used a customized ARTEX deployment alongside AI coding assistants, automation scripts, and specialized agent instructions. Rather than conducting each stage manually in an isolated session, the operator created a workflow that could divide tasks among AI-assisted workers, preserve prior findings, and resume work across different accounts and servers.
The recovered records indicated a focus on obtaining personal and financial information. Requests included collecting user records, retrieving customer and transaction data, accessing administrative functionality, and examining ways to obtain payment-related credentials or exploit application workflows.
The infrastructure also included proxy routing, SSH tunnels, and cloud backups for investigation files. These components helped maintain the operator's working environment and organize information collected during activities against different targets.
Zscaler assessed the campaign as likely financially motivated. Its findings also pointed to a Chinese-speaking operator, based on language preferences, instructions, time-zone settings, and references in the recovered records. These observations should be understood as the researchers' attribution assessment rather than definitive proof of the operator's identity.
2. Technical Description
The campaign combined several components into a coordinated attack workflow.
2.1 AI-Assisted Penetration-Testing Framework
The operator deployed ARTEX, an AI-assisted penetration-testing framework using a planner-and-worker architecture. A high-level objective could be divided into smaller tasks for specialized agents to execute and document.
The customized deployment incorporated additional tools for web, network, and Android analysis, including:
nmapsqlmapnucleihttpxjadxapktoolfrida-tools
These tools have legitimate security-testing uses. Their presence in the recovered environment is significant because they were integrated into a workflow used to investigate targets without authorization, according to the reported activity.
2.2 Multiple AI Assistants and Model Routing
The operator configured several AI-assisted workspaces to use different model backends. The recovered configurations referenced DeepSeek, GLM, and Grok, alongside applications such as Claude Code and Codex.
The setup used API-compatible gateways and a bridge service to connect some client applications to alternative model services. Custom instructions assigned distinct roles to different AI sessions.
Importantly, the model identifiers recorded in the configuration files were requested identifiers. Zscaler noted that the underlying model serving each session could not be independently verified.
2.3 Automated Context and State Management
The operator created scripts and project files to preserve investigation progress across AI sessions. These records included target scopes, discovered attack surfaces, completed tasks, notes, findings, and supporting evidence.
A context hook was configured to read project files when a new prompt arrived and provide the AI session with a summary of prior work. Specialized agent instructions supported tasks such as API discovery, source-code review, and Android application analysis.
This arrangement reduced the need to restart investigations from scratch whenever a new session or agent was used.
2.4 Network Routing and Proxy Infrastructure
The operation used a work server and a jump server. The jump server acted as the default route toward targets and hosted additional AI sessions.
Proxy configurations supplied separate outbound paths for individual projects. SSH tunnels supported communication between the servers, while monitoring tasks were configured to restart supporting services if interrupted.
The operator also used cloud storage to archive reports, source code, collected data, evidence, and backups.
2.5 Targeting and Data Collection
Recovered prompts showed requests to investigate websites, APIs, mobile applications, and other internet-facing services.
The desired outcomes included collecting personal information, accessing administrative functions, obtaining transaction records, retrieving payment-related credentials, and examining potential weaknesses in application workflows.
The investigation records also included requests to process extracted information, decrypt encrypted fields, interpret database schemas, and prepare datasets for possible monetization.
3. Attack Overview
The following sequence summarizes the operational workflow described by Zscaler. It is a high-level reconstruction of the recovered environment, not a claim that every target experienced the same steps.
Stage 1: Infrastructure Preparation
The operator established a work server and a jump server, deployed a customized ARTEX environment, configured AI services, and prepared proxy routing.
Stage 2: AI Workflow Configuration
Different AI sessions were assigned roles such as initial analysis, collection, investigation, and peer review. Shared instructions and synchronization scripts maintained consistency across accounts.
Stage 3: Target Selection
The operator maintained project records describing selected websites, servers, applications, and their discovered attack surfaces.
The recovered prompt history indicated targeting across several industries and countries rather than a single organization or sector.
Stage 4: Automated Reconnaissance
ARTEX and supporting tools were configured to perform tasks related to web, network, and Android analysis. AI agents were instructed to break broader objectives into smaller tasks.
Stage 5: Findings and Peer Review
The operator instructed one AI session to review findings generated by another. For example, GLM was assigned review work involving results from DeepSeek and its subagents.
The purpose was to identify missed attack surfaces, review previous results, and continue analysis.
Stage 6: Unauthorized Access and Data Acquisition Objectives
Prompt histories included requests to obtain user records, personal information, transaction histories, payment-related credentials, and administrative access.
These records support the assessment that the activity went beyond routine reconnaissance and included objectives involving unauthorized data access.
Stage 7: Data Processing and Storage
The operator organized findings, reports, evidence, and data exports into separate directories and configured backups to cloud storage.
Requests to structure, translate, and filter extracted data suggested that the collected information was being prepared for further use and potential monetization.
4. Technical Analysis
4.1 Work Server and Jump Server
Zscaler identified two primary roles within the operator's infrastructure.
The work server stored project files, scripts, AI settings, and the local ARTEX deployment. It served as the central environment for organizing tasks and maintaining investigation records.
The jump server, identified by the IP address 38.244.50.120, acted as the default route for connections to targets. It also hosted additional AI sessions, including sessions assigned research and review tasks.
The two-server arrangement separated portions of the operator's workflow while allowing files and instructions to be synchronized.
Defensive significance: Infrastructure that coordinates automated reconnaissance may generate repeated outbound connections, unusual proxy activity, and patterns of access across many unrelated targets. Network monitoring and correlation can help identify such activity.
4.2 Custom Synchronization Scripts
Zscaler identified three scripts supporting the operator's workflow:
| Script | Reported purpose |
|---|---|
tools/sync-all.sh | Synchronizes AI instructions, task guides, saved notes, and tools across accounts on the jump server |
sync-memory.sh | Copies saved AI notes and memory files to locations on the jump server |
tools/sync-cc-users.sh | Copies shared Claude Code instructions and supporting files to designated accounts on the work server |
The scripts also accounted for differences between destination accounts and model configurations. Backup and validation checks were included to reduce the risk of overwriting modified files.
This indicates that the operator intended to maintain a consistent working environment across multiple sessions and accounts.
4.3 Customized ARTEX Deployment
ARTEX uses a planner-and-worker model in which an operator defines an objective and AI-assisted agents divide the work into smaller tasks.
The recovered Docker Compose configuration defined an ARTEX container and a PostgreSQL 16 database. The configuration included:
ARTEX interface port:
8890PostgreSQL mapped to:
127.0.0.1:5435Custom container image:
artex-local:perf2-20261003
The corresponding Dockerfile added tools including nmap, sqlmap, nuclei, httpx, jadx, apktool, and frida-tools, along with Python libraries used for web processing and spreadsheet output.
These details describe the recovered environment. They should not be interpreted as universal default settings for ARTEX or as indicators that every ARTEX deployment is malicious.
4.4 AI Model Setup and Delegation
The recovered configuration files showed multiple requested model identifiers.
| Workspace or component | Requested model identifier | Reported role |
|---|---|---|
| Local ARTEX deployment | deepseek-v4.1-flash | Backend for the ARTEX workflow |
| Codex configuration | deepseek-v4-pro | Alternative coding-agent interface |
Local account named gemini, running Claude Code | deepseek-v4.1-flash | Workspace using shared instructions |
Local account named glm, running Claude Code | glm-5.3 | Workspace configured for GLM |
Jump-host accounts root and claude | glm-5.3, glm-5.3-flash | Assigned analysis and collection roles |
Jump-host account grok | grok-4.7-xhigh, grok-4.7-build-fast | Research and analysis roles |
Jump-host account named gemini | xcai/deepseek-v4.1-flash | Workspace configured to use DeepSeek |
The operator also installed a bridge service called grok2api to support a Claude Code integration configured to use Grok as its underlying model.
The configuration records establish that the operator requested these model identifiers. They do not independently prove which underlying model responded to every request.
4.5 Persistent Project State
The operator created separate folders for individual investigations. These files helped AI sessions retain the context needed to continue work across sessions.
| File or folder | Purpose |
|---|---|
scope.md | Records the websites, servers, or applications selected for investigation |
surface.md | Records discovered pages, APIs, and other entry points and their testing status |
notes.md | Stores completed work, decisions, and next steps |
findings/ | Contains vulnerability reports and findings |
evidence/ | Stores supporting files, captured responses, and tool output |
A context hook named tools/project_context_hook.py was configured to read project records when a new prompt arrived and attach a summary of progress and relevant files.
Specialized instructions were also used to assign tasks to agents responsible for API discovery, source-code analysis, or Android application analysis. The primary AI session was instructed to review agent outputs before recording findings.
The recovered instructions also told the AI systems to assume that supplied targets were authorized and reduced the need for operator approval in some configurations. This is a significant safety and security concern because it removes an important checkpoint before automated actions are taken against external systems.
4.6 Proxy Exits, Tunnels, and Cloud Backup
The operator instructed AI tools to route target connections through the jump server. Additional proxy configurations provided separate outbound IP addresses for individual projects.
SSH tunnels enabled communication between the work server and jump server. Background tasks were configured to restart supporting services and tunnels if they stopped.
The operator also archived investigation material to Backblaze B2. Reports, evidence, source code, data exports, and backups were organized into separate directories.
This infrastructure helped maintain continuity and organize data collected across multiple targets.
Defensive significance: Organizations should monitor suspicious proxy activity, unexpected SSH tunnels, repeated connections to unfamiliar infrastructure, and unauthorized transfers to cloud-storage services.
4.7 Global Targeting Across Multiple Industries
The prompt history extended beyond South Korean financial institutions. It referenced targets in numerous industries and countries.
Representative target categories included:
| Target category | Objectives recorded in the prompt history |
|---|---|
| Social and dating services | Collect frontend files and subdomains; seek access to user records |
| SMS providers | Seek administrative access, SMS-channel credentials, or unauthorized messaging capability |
| Retail platforms | Obtain customer order data |
| Educational services | Obtain course-purchase records or user information |
| Investment services | Review prior investigation records and obtain current data |
| Cryptocurrency and digital assets | Find weaknesses in NFT-related workflows |
| Legal claims administration | Review claims workflows and examine opportunities for additional access |
The prompt history referenced organizations or targets associated with Canada, China, Cuba, Hong Kong, India, Indonesia, Japan, Kazakhstan, Kenya, Malaysia, Nigeria, South Africa, South Korea, Taiwan, Thailand, the United Arab Emirates, the United Kingdom, the United States, and Vietnam.
This geographic spread suggests opportunistic targeting across multiple regions rather than an operation limited to a single national sector.
4.8 Personal Information and Financial Data Collection
The recovered prompts indicated an explicit interest in structured personally identifiable information (PII), mobile phone numbers, and financial data.
In one example, the operator expressed dissatisfaction with database extracts that lacked phone-number fields and requested alternative injection methods. Another prompt questioned why a SQL query targeting demographic information for women over 30 returned only approximately 300 records.
The operator also directed AI models to:
Decrypt encrypted fields.
Parse transaction histories.
Translate database schemas into Chinese.
Remove irrelevant metadata.
Search for payment gateway API keys.
Examine ways to obtain unauthorized administrative access.
Prompts asking where compromised Korean data was commonly sold on underground marketplaces further supported the researchers' assessment that the activity was financially motivated.
Zscaler also reported instructions warning AI models not to initiate direct account withdrawals because unauthorized transactions could trigger alerts or create traceable records.
These findings suggest that the operator sought data and access that could be exploited or monetized, rather than limiting activity to vulnerability discovery.
4.9 Threat Attribution
Zscaler's recovered records pointed to a Chinese-speaking operator. The supporting observations included:
Prompts written predominantly in Chinese.
Instructions to use Chinese as the default response language.
References to China Standard Time.
Requests to translate exported data fields into Chinese.
References to Chinese services as domestic services.
A mainland China-format phone number supplied for accounts associated with a proxy provider and cloud-phone service.
These observations informed Zscaler's assessment. They do not independently establish the operator's identity, location, or state affiliation.
4.10 Why AI Assistance Matters in This Campaign
The significance of this campaign lies in how AI was integrated into the operator's workflow.
The recovered environment supported:
Breaking broader objectives into smaller tasks.
Assigning work to specialized agents.
Reviewing one agent's findings with another.
Maintaining project state across sessions.
Synchronizing instructions and supporting files.
Continuing investigations across multiple accounts and servers.
Organizing evidence and data for later analysis.
AI assistance does not eliminate the need for an operator or automatically guarantee successful exploitation. In this case, the records showed an operator directing the workflow, choosing targets, assigning objectives, and configuring the supporting infrastructure.
The key defensive lesson is that AI-assisted activity may appear as a combination of ordinary security tools, coding-agent sessions, automation scripts, and conventional network infrastructure. Effective detection therefore requires correlating behaviors and their context instead of relying on the presence of one tool or AI model.
5. Indicators of Compromise (IOCs)
Zscaler published the following network indicators associated with the investigated campaign. IP addresses and ports are reproduced from the source.
5.1 Network Indicators
| Indicator | Description |
|---|---|
23.248.226.250 | Work server |
23.248.226.251 | Work server |
23.248.226.252 | Work server |
23.248.226.253 | Work server |
23.248.226.254 | Work server |
38.244.50.120 | Jump server |
154.201.79.246:443 | Proxy pool API |
205.214.59.31:443 | Upstream SOCKS proxy |
161.8.161.140:11005 | Upstream SOCKS proxy |
IOC handling guidance: These indicators should be treated as investigation leads. Validate them against current threat intelligence, organizational logs, and the original report before blocking them. An IP address alone does not prove that a particular connection was malicious.
5.2 Additional Artifact and Detection Context
The article identifies the following infrastructure and artifacts as relevant to the investigation:
Customized ARTEX deployment.
tools/sync-all.shsync-memory.shtools/sync-cc-users.shtools/project_context_hook.pygrok2apixcai-shim.pyscope.mdsurface.mdnotes.mdfindings/evidence/
These names describe files, scripts, and components discussed in the report. They are not necessarily unique malware indicators and should be investigated in context.
Zscaler also reported its own detection name for the campaign: AI.HackTool.ARTEX.
5.3 IOC Investigation Workflow
Security teams can use the indicators to support threat hunting:
Search firewall, DNS, proxy, and network-flow logs for the listed IP addresses and ports.
Determine whether connections originated from approved systems or unexpected hosts.
Investigate unusual SSH tunnels, SOCKS proxy activity, and connections to unfamiliar proxy services.
Look for unauthorized use of penetration-testing utilities and automated scanning against external targets.
Review endpoint and server logs for suspicious scripts that synchronize project files or maintain AI-session context.
Examine unexpected cloud-storage uploads and transfers of reports, data exports, or source code.
Preserve relevant logs and artifacts before making changes that could destroy evidence.
Validate matches and record confirmed findings separately from unverified leads.
6. Potential Impact
6.1 Personal Information Exposure
The campaign's prompt history indicated an interest in obtaining structured personal information, mobile phone numbers, and user records. Unauthorized access to this information can lead to privacy violations, fraud, targeted scams, and further account compromise.
6.2 Financial Data and Credential Theft
The recorded objectives included transaction histories, payment gateway API keys, and access to financial or cryptocurrency-related services. Such information can expose organizations and their customers to fraud or other financial abuse.
6.3 Unauthorized Administrative Access
Attempts to obtain administrative privileges can give an attacker broader control over applications, databases, or infrastructure. The extent of the damage depends on the access obtained and the protections surrounding the targeted environment.
6.4 Scaled Reconnaissance
The use of AI-assisted task delegation and persistent project records can help an operator organize activity across numerous targets. This may increase the breadth of reconnaissance and the number of organizations exposed to unauthorized testing.
6.5 Operational Disruption
Unauthorized scanning, exploitation attempts, and misuse of application workflows can affect service availability, trigger security incidents, or require organizations to suspend affected functionality while investigations are conducted.
6.6 Regulatory and Reputational Consequences
Exposure of personal or financial information may lead to regulatory obligations, incident-response costs, customer notification requirements, and reputational harm, depending on the affected organization and applicable laws.
7. Recommendations
7.1 Immediate Actions
1. Review the reported network indicators
Search network telemetry for the listed work servers, jump server, and proxy infrastructure. Confirm whether any connections are associated with legitimate business activity or unauthorized reconnaissance.
2. Investigate suspicious scanning activity
Review web application, firewall, WAF, and endpoint logs for unusual patterns involving nmap, sqlmap, nuclei, httpx, or other reconnaissance tools. These utilities have legitimate uses, so evaluate them in context.
3. Protect exposed applications and APIs
Identify internet-facing applications, APIs, and administrative interfaces. Review access controls, authentication requirements, rate limits, and authorization checks.
4. Examine suspicious data access
Investigate unusual database queries, bulk exports, unexpected access to customer records, and requests involving personal information or transaction histories.
5. Review outbound connections
Look for unauthorized SSH tunnels, SOCKS proxy use, suspicious proxy APIs, and unexpected cloud-storage transfers.
6. Audit privileged access
Review accounts with administrative privileges, unexpected privilege changes, service-account permissions, and access to sensitive application or database infrastructure.
7. Preserve evidence
Retain relevant server logs, network records, application requests, database audit events, endpoint telemetry, and suspicious scripts.
8. Contain confirmed incidents
If unauthorized access is confirmed, restrict compromised accounts and sessions, isolate affected systems where appropriate, remediate the entry point, and investigate whether sensitive information was accessed or transferred.
7.2 Preventive Actions
A. Zero Trust and Exposure Management
Minimize unnecessary public exposure of applications, APIs, and administrative interfaces.
Enforce least-privilege access for users, services, and administrators.
Require strong authentication for sensitive applications.
Review internet-facing assets regularly.
Prioritize remediation of exposed services and confirmed vulnerabilities.
Segment sensitive workloads from general user and internet-facing networks.
B. Web Application and API Security
Validate authorization on every sensitive API operation.
Use parameterized database queries and appropriate input validation.
Apply rate limiting and abuse detection to sensitive endpoints.
Monitor bulk data access and unusual query patterns.
Test authentication, authorization, and data-export workflows.
Protect payment, customer, and identity systems with additional controls.
C. Identity and Privileged Access
Enforce MFA for administrative and remote access.
Apply least privilege to user, service, and application accounts.
Audit privileged access and remove unnecessary permissions.
Rotate credentials that may have been exposed.
Monitor unusual account creation and privilege changes.
Use controlled, time-limited administrative access where feasible.
D. Network Monitoring
Monitor unusual scanning and probing activity.
Investigate unexpected proxy services, SSH tunnels, and outbound connections.
Correlate network events with endpoint processes and application logs.
Restrict outbound traffic according to business requirements.
Alert on suspicious transfers to external storage and data-hosting services.
E. Data Protection
Classify sensitive personal, financial, and transaction data.
Restrict bulk exports to authorized roles and business purposes.
Maintain database audit logging for sensitive tables and operations.
Encrypt sensitive data at rest and in transit.
Monitor unusual access to payment-related credentials and API keys.
Establish clear procedures for responding to suspected data theft.
F. AI and Automation Governance
Define approved uses of AI coding assistants and automated security-testing frameworks.
Require documented authorization and scope for security testing against external assets.
Use approval gates for high-impact actions involving production systems or sensitive data.
Monitor automation that launches scans or executes commands across multiple targets.
Restrict the credentials and permissions available to automated agents.
Preserve audit records showing the human operator, task scope, tools used, and actions performed.
Review AI-generated findings before treating them as confirmed vulnerabilities.
Prevent unapproved agents from accessing sensitive data or transferring it to external services.
These measures address the risks demonstrated by the campaign without assuming that all AI-assisted security tools are inherently malicious.
8. Incident Response
If an organization identifies activity consistent with the reported campaign, it should follow a coordinated response process.
Step 1: Validate the Alert
Confirm whether the event involves a reported network indicator, suspicious reconnaissance, unauthorized data access, or unusual administrative activity.
Step 2: Establish Scope
Identify affected hosts, accounts, applications, APIs, databases, and network segments. Determine whether the activity was limited to scanning or involved successful access.
Step 3: Preserve Evidence
Collect relevant network logs, application requests, database audit records, endpoint telemetry, and suspicious scripts. Preserve timestamps and relevant system metadata.
Step 4: Contain the Activity
Restrict unauthorized accounts, sessions, or network paths. Isolate affected systems where necessary and avoid disrupting evidence collection.
Step 5: Investigate Data Exposure
Determine whether personal information, financial records, credentials, API keys, or other sensitive data were accessed or exported.
Step 6: Remediate
Fix exploited vulnerabilities, correct authorization weaknesses, remove unauthorized access, rotate exposed credentials, and strengthen affected controls.
Step 7: Recover and Monitor
Restore affected services safely, validate security controls, and monitor for renewed access or similar behavior.
Step 8: Document Lessons Learned
Record the attack timeline, confirmed findings, response actions, and improvements needed in exposure management, identity security, monitoring, and automation governance.
9. MITRE ATLAS Framework Mapping
Zscaler explicitly included the following AI-focused mappings in its report. These are reproduced as source-reported mappings.
| ID | Technique | Relevance described in the report |
|---|---|---|
AML.T0040 | AI Agent Tool Integration | AI assistants integrated with execution-capable tools |
AML.T0043 | Autonomous AI Reconnaissance & Exploitation | Planner/worker agents divide objectives into executable tasks |
AML.T0048 | Multi-Model LLM Routing & Delegation | Different model backends assigned to different workspaces |
AML.T0051 | LLM Peer Review & Verification | One AI session reviews another session's findings |
AML.T0054 | AI Guardrail Bypass / System Prompt Engineering | Instructions tell AI sessions to assume targets are authorized and reduce approval requirements |
AML.T0052 | Persistent AI Context & State Management | Hooks and synchronization scripts preserve project context across sessions |
These mappings follow the Zscaler article. They should not be silently substituted with a different taxonomy when the purpose is to reproduce the source's findings.
10. Conclusion
The global AI-assisted hacking campaign described by Zscaler shows how an operator can combine a penetration-testing framework, AI coding assistants, specialized agents, custom scripts, proxy infrastructure, and persistent project records into a coordinated workflow.
The recovered prompts indicate that the activity extended across multiple industries and countries, with objectives involving personal information, financial data, payment-related credentials, and unauthorized administrative access. Zscaler assessed that the activity was likely financially motivated.
The investigation also demonstrates that AI assistance does not operate in isolation. Its effectiveness in this case depended on conventional infrastructure, security-testing tools, network routing, automation scripts, and the operator's direction.
Organizations should focus on the behaviors that enable these operations: unauthorized reconnaissance, application and API abuse, excessive data access, suspicious proxy activity, and privilege escalation. Strong identity controls, secure application design, continuous exposure management, network monitoring, and governance of automated testing can help reduce the risk.
