1. Description
Artificial intelligence is changing how cybercriminals plan, automate, and execute attacks. However, not every attack involving AI is fully autonomous, and the extent of AI involvement varies considerably between operations.
In its report published on October 27, 2026, TrendAI examined three separate attack infrastructures associated with operations targeting organizations in Japan and other regions. Researchers reconstructed activity from files left in publicly accessible directories on attacker-controlled servers.
The investigation identified three distinct patterns:
AI-driven execution: An AI-agent platform called CyberStrikeAI coordinated multiple stages of an attack workflow.
AI-assisted operations: Human operators used large language models (LLMs) to accelerate planning and tool development while retaining control over attack decisions.
Limited AI-assisted development: Experienced operators appeared to use AI primarily to develop scripts and tools, while conducting the broader operation themselves.
The report followed a series of data leaks affecting Japanese organizations during autumn 2026 and a cybersecurity alert issued by Japan's National Cybersecurity Office on October 9.
TrendAI emphasized that the three infrastructures examined were distinct from those discussed in earlier reporting about ARTEX, an autonomous AI penetration-testing system. The findings illustrate how attackers can use AI in different ways without necessarily relying on new vulnerabilities.
2. Technical Description
The investigation focuses on the use of AI across reconnaissance, vulnerability discovery, exploitation, command and control (C2), persistence, data exfiltration, and reporting.
AI-driven attack automation
In the first case, CyberStrikeAI was used to coordinate an attack workflow through multiple AI agents. The exposed management console contained evidence of agents performing different tasks, including reconnaissance, vulnerability assessment, exploitation, and data handling.
The platform generated structured reports and organized results in a repeatable format. Researchers confirmed that AI was used to drive the broader workflow rather than merely assist an operator with isolated tasks.
AI-assisted human operations
The second case involved an operator who used an LLM to develop attack plans and tools. The human operator remained responsible for decisions and execution, while AI accelerated selected tasks.
The investigation also found evidence of attempts to obtain access to external AI services and GPU infrastructure, suggesting that AI resources themselves can become targets.
AI-assisted tool development
In the third case, the operator appeared to use AI for limited tasks, particularly script and tool development. Researchers identified multiple stylistic indicators consistent with LLM-generated material, but did not recover direct evidence such as prompts or API keys.
TrendAI therefore classified this as a high-confidence assessment of AI involvement, rather than confirmed use.
These distinctions matter because evidence that an attacker used AI to write a script does not establish that AI independently planned or executed the entire attack.
3. Attack Overview
The three cases demonstrate different ways AI can support cyberattack operations.
Reconnaissance: Attackers identify publicly accessible systems, legacy web applications, VPN devices, and exposed management interfaces.
Target prioritization: Collected information is organized into target lists and assessment records.
Tool and plan generation: AI assists with developing scripts, preparing attack plans, and organizing technical workflows.
Automated or human-directed exploitation: Depending on the case, an AI-agent platform or a human operator attempts to exploit accessible weaknesses.
Access establishment: Successful activity may provide access to applications, databases, administrative interfaces, or internal networks.
Post-compromise activity: Depending on the observed case, attackers may deploy web shells, restore backdoors, access sensitive information, or attempt data theft.
Evidence collection and reporting: Results are stored in structured files, logs, and reports, helping operators track activity across multiple targets.
Not every stage occurred in all three cases. TrendAI specifically noted that many attempts did not result in deep compromise, and attacker-generated reports sometimes overstated success.
4. Technical Analysis
4.1 Case A: Autonomous execution through CyberStrikeAI
The first infrastructure provided the clearest evidence of AI-driven attack execution.
Researchers confirmed that CyberStrikeAI, an open-source AI-agent penetration-testing platform, was used to coordinate reconnaissance, vulnerability discovery and validation, exploitation, C2, persistence, data exfiltration, and reporting.
The system divided tasks among specialized agents. Records included a configuration involving 44 agents, with different agents handling activities such as credential searches, reconnaissance, and exploitation.
The platform produced structured reports and evidence indexes for at least 19 identifiable organizations, including at least 16 in Japan. Taiwan and Hong Kong also appeared among the targets.
Despite the scale of the activity, evidence of deep compromise was limited to a small proportion of targets. The platform recorded unsuccessful outcomes for many systems, including cases without an established session, web shell, or data exfiltration.
The clearest intrusion chain began at an unauthenticated, publicly accessible administration interface on a legacy web system. Researchers did not identify a novel vulnerability as the entry point in that case.
Key finding: AI enabled parallel operations, repeatability, and scale, but the attack still depended on an accessible weakness.
4.2 Case B: Human-led attacks supported by AI
The second infrastructure demonstrated a different operating model. A human attacker used an LLM to accelerate planning and tool development while retaining control over the operation.
Researchers recovered an LLM-generated attack plan that covered reconnaissance and attempts against services providing AI and GPU resources. They also identified evidence that the operator presented requests as legitimate authorized assessments to circumvent the model's safety restrictions.
The investigation found that the operator targeted legacy web systems and attempted to access external AI infrastructure, including workflow platforms, preprocessing systems, and inference services.
In one reported intrusion, the attacker obtained extensive control involving FTP, a database, administrator privileges, a web shell, and SSH. The operator repeatedly restored backdoors after cleanup. In another case, a geographic information system exposed through default credentials provided access to large volumes of cadastral and commercial records, Windows credentials, and email.
TrendAI confirmed data theft from the recovered material and download records on the attacker-controlled infrastructure.
However, the researchers found no evidence that AI autonomously executed the attacks in this case. The operator primarily relied on existing tools, handwritten scripts, and personal judgment.
Key finding: AI can improve the efficiency of a human-led operation, while external AI services and GPU environments can themselves become attack targets.
4.3 Case C: AI assistance in tool development
The third infrastructure showed signs of a more limited use of AI. Researchers identified multiple stylistic indicators consistent with LLM-generated scripts and command sequences, including carefully structured comments and section headings.
The operator maintained organized directories, logs, evidence, success lists, and verification mechanisms designed to reduce false positives.
TrendAI identified at least six campaigns during 2026, with several operating in parallel. Targets included exposed VPN devices and Oracle WebLogic servers.
The targeted edge devices included Cisco SD-WAN, Citrix NetScaler, F5 BIG-IP, and Yamaha RTX1210. The observed activity involved attempts to abuse published vulnerabilities or management functions on some products.
For Yamaha RTX1210 devices, the investigation found that attackers used credentials retained in obtained configuration files rather than exploiting a product vulnerability. Those credentials enabled direct connections to corporate internal networks.
The evidence supporting AI use in this case was based on multiple stylistic indicators. Researchers did not find direct prompts or API keys, so AI involvement was assessed with high confidence rather than classified as definitively confirmed.
Key finding: AI may contribute to an attack without controlling its overall workflow, and configuration weaknesses can remain exploitable even when no new vulnerability is involved.
4.4 What AI changes—and what it does not
TrendAI's comparison highlights several important differences between traditional attack operations and those enhanced by AI.
What AI can change:
The number of targets handled in parallel.
The speed and repeatability of reconnaissance and exploitation attempts.
The generation of scripts, plans, and structured reports.
The adaptation of attack materials to a target's language.
The ability to organize and process large volumes of results.
What remains largely unchanged:
The importance of exposed services and vulnerable applications.
The risks associated with default credentials and weak access controls.
The need to establish a viable initial access path.
The difference between an attempted attack and a confirmed compromise.
The report also identified limitations in AI-generated outputs, including overly optimistic success assessments and the possibility of planned actions being recorded as completed.
These limitations make independent verification important for both attackers and defenders.
4.5 Attribution and evidence limitations
TrendAI did not establish whether the three infrastructures shared a common operator. Simplified Chinese was used across the three environments, but the researchers cautioned that working language alone cannot establish the operators' location, identity, or affiliation.
The investigation relied on attacker-side files and did not include victim-side logs, disk images, or network records. As a result, the findings cannot independently establish the complete extent of damage for every target.
The report also distinguishes confirmed AI use from high-confidence inferences based on coding style and other indirect indicators.
These limitations should be considered when interpreting both the scale of the activity and its attribution.
5. Indicators of Compromise (IOCs)
No IOCs are included in this article.
TrendAI explicitly withheld indicators from its report to avoid interfering with ongoing investigations and notifications to potentially affected organizations.
The absence of published IOCs does not mean the activity is harmless. Organizations should instead use the reported behaviors, target types, and defensive recommendations to guide threat hunting and security reviews.
6. Potential Impact
AI-assisted and AI-driven attacks can create significant operational risks for organizations:
Increased attack scale: Automated workflows can assess multiple targets concurrently.
Faster exploitation attempts: AI-assisted planning and tool development can shorten the time needed to prepare attacks.
Data exposure: Successful access to web systems, databases, and internal networks can expose sensitive business and personal information.
Persistent unauthorized access: Attackers may deploy or restore web shells and backdoors when security weaknesses remain unaddressed.
Compromise of AI infrastructure: Exposed inference services, AI workflows, and GPU resources can become targets for unauthorized access or resource abuse.
Reduced detection effectiveness: Distributed, low-frequency attempts can be difficult to recognize when organizations investigate alerts in isolation.
Expanded attack surface: Legacy applications, exposed VPN devices, misconfigured services, and retained credentials provide opportunities for attackers.
These are risks supported by the behaviors and findings discussed in the report. The investigation does not establish that every targeted organization suffered compromise.
7. Recommendations
Immediate Actions
Inventory internet-facing assets: Identify public web applications, legacy systems, administration interfaces, VPN devices, and exposed middleware.
Review access controls: Check for default credentials, exposed management interfaces, excessive API permissions, and unnecessary public access.
Investigate distributed scanning: Correlate low-frequency connection attempts across web, VPN, firewall, and endpoint logs.
Audit VPN and edge-device configurations: Review configuration backups for retained credentials and investigate unexpected access to internal networks.
Secure AI infrastructure: Identify externally exposed inference endpoints, AI workflow platforms, GPU environments, and related management interfaces.
Check for post-compromise artifacts: Investigate unexpected web shells, restored backdoors, suspicious administrative activity, and unusual data transfers.
Verify reported outcomes: Do not classify an attack as successful solely because an attacker-side report claims success. Validate findings using independent telemetry.
Preventive Measures
Patch and isolate legacy systems: Keep public-facing platforms updated and remove or restrict systems that can no longer be securely maintained.
Remove unnecessary exposure: Restrict administrative interfaces and internal business applications to authorized networks and identities.
Eliminate default and retained credentials: Rotate exposed credentials, protect configuration backups, and enforce strong authentication.
Apply least privilege and segmentation: Limit web-server access to internal systems and prevent compromised applications from reaching unnecessary resources.
Monitor low-and-slow activity: Use cross-system correlation and behavioral analytics to detect distributed scanning that may not trigger individual alert thresholds.
Strengthen AI governance: Define authorized use of AI tools, review sensitive data handling, and monitor access to internal AI services.
Validate security testing results: Treat automated reports as findings requiring verification, especially when AI-generated assessments claim successful exploitation.
Prepare for rapid vulnerability response: Maintain an effective patching process and use temporary mitigations when permanent updates cannot be applied immediately.
8. Conclusion
TrendAI's investigation demonstrates that AI is being incorporated into cyberattacks at several levels, from fully automated attack workflows to human-directed operations and limited assistance with script development.
The most important finding is not simply that attackers are using AI, but that AI can increase the scale, speed, and repeatability of operations against systems that remain exposed or poorly configured.
The cases also show why evidence must be interpreted carefully: large-scale scanning does not necessarily mean widespread compromise, stylistic clues do not always prove AI use, and attacker-generated reports may overstate success.
For defenders, the priorities remain clear: reduce exposed attack surfaces, eliminate default credentials, secure VPN and AI infrastructure, monitor distributed activity, and independently validate suspected compromises. AI may change how quickly attackers find weaknesses, but consistent security fundamentals remain essential to preventing them from succeeding.
