TrendAI researchers report, in Japanese, on three attack infrastructures they reconstructed from files the operators left in open directories, each focused on Japan and separate from the ARTEX activity GMO Cybersecurity by Ierae and CrowdStrike reported. They class one as AI-driven, running the open-source CyberStrikeAI tool with 44 agents against at least 19 organizations in October 2026; one as AI-assisted, with a human operator using a language model for planning and tooling in August and September 2026; and one as showing limited AI help with tool writing, rated high confidence from stylistic traces only. TrendAI saw only attacker-side files and could not verify outcomes from victims’ systems, the full damage, operator identities or models used. It counts success as confirmed only where attacker files provide direct evidence. The entry points it describes were known weaknesses such as exposed legacy systems or default settings.
It sets three operations side by side and grades how much of each AI carried and how sure the evidence is, separating attempted from confirmed compromise, rather than treating AI-enabled attacks as one category.
Key facts
As stated in the sources, with where to find them.
- TrendAI analyzed only files on servers the attackers had left publicly exposed (open directories), with no victim-side logs, disks or traffic. It says this shows what the attackers prepared, tried and recorded, but not whether individual attempts succeeded, the scale of damage, or who the attackers are. It treats the attackers' many self-reported successes as confirmed only where attacker-side files hold direct evidence, and grades claims as confirmed (direct evidence in attacker files), high confidence (several indirect signs agree) or inferred.TrendAI report, section on what the report does and does not show
- TrendAI says all three infrastructures are separate from the ARTEX reports by GMO Cybersecurity by Ierae and CrowdStrike, that it observes many open-directory infrastructures using AI to attack many countries indiscriminately, often including Japan, and that it chose three with a strong Japan focus. Activity in all three fell in 2026, much of it before the GMO and CrowdStrike reports. It withholds indicators of compromise, plans to give them to law enforcement, and is preparing to notify suspected victims.TrendAI report, introduction and overview
- Case A (AI-driven, confirmed): the operator abused CyberStrikeAI, a public open-source AI-agent penetration-testing tool whose exposed management console is in Simplified Chinese. TrendAI describes a repeatable workflow from reconnaissance through validation, exploitation, command and control, persistence, exfiltration and automatic report writing, split among 44 agents (one record reads, in TrendAI's words, credential-search agent 8/44), with Japanese-language forms and fake login pages generated by the platform. TrendAI does not describe the platform's internals, C2 or persistence methods.TrendAI report, Case A
- Case A scale and outcome: in a short burst in October 2026, at least 19 organizations, at least 16 of them in Japan, with targets also in Taiwan and Hong Kong, were subjected to validation activity, each leaving a same-format report and evidence index; contact with each target was kept infrequent, which TrendAI says was to avoid detection. For many targets the platform itself recorded no session, no web shell and no exfiltration; deep compromise holds for only a very small share, and the clearest intrusion chain, at one company, began at an unauthenticated public admin page on a legacy web system. TrendAI calls the case AI-driven rather than AI-assisted because of its scale and consistency.TrendAI report, overview table and Case A
- Case B (AI-assisted, confirmed): a human-led operator over about 24 days in August and September 2026 left LLM output files with attack plans against AI and GPU services, and had told the model the work was an authorized assessment, a false premise TrendAI says was meant to get around the model's safeguards. Its target lists included AI workflow platforms, LLM preprocessing and inference services, and it built tools to hijack and pool others' AI service access, which TrendAI infers was to obtain inference capacity cheaply. It scanned more than 6,000 IP addresses in Japan and focused on a few companies.TrendAI report, overview table and Case B
- Case B outcomes: TrendAI confirms from attacker-side files and download records full control of at least one company's systems, with backdoors restored after cleanup, and the theft of hundreds of thousands of cadastral and commercial records, Windows credentials and email from another company through a GIS server left with default credentials. It found no evidence that AI carried out attacks autonomously; the operator relied mainly on hand-written scripts, off-the-shelf tools and own judgment, and hundreds of attempts to break into an AI/GPU cluster all ended in errors.TrendAI report, Case B
- Case C (limited AI help, high confidence, not confirmed): an experienced operator ran at least six campaigns from January to October 2026, with campaigns running in parallel especially from September, against exposed VPN devices and WebLogic servers, including Cisco SD-WAN, Citrix NetScaler, F5 BIG-IP and Yamaha RTX1210 devices; for the RTX1210, credentials left in retrieved configuration files gave direct connections into corporate networks. Lists held more than 1,300 WebLogic candidates and more than 35 VPN entry points; many VPN connections succeeded but deep compromise was limited. TrendAI bases the AI rating on stylistic traces in command histories and scripts, found no prompts or API keys, and says a skilled person writing in that style cannot be ruled out.TrendAI report, overview table and Case C
- TrendAI says all three operators worked in Simplified Chinese, which it calls only a weak clue to location or affiliation, and that it does not know whether the three share an operator and has no evidence linking them to the ARTEX actor GMO or CrowdStrike reported. All three picked targets opportunistically across sectors, unlike the country- and sector-focused campaign GMO described. TrendAI calls its counts lower bounds.TrendAI report, comparison of the three cases
- TrendAI lists as unresolved the attackers' identity, whether the cases share an operator, the real scale of damage, and which AI models were used. It says AI changed the number of targets handled in parallel, speed, target-language output and automated record-keeping, but not the kinds of vulnerabilities exploited, the kinds of entry point or the rate of deep compromise, and that the attacker-side AI recorded planned steps as done and rated its own success too optimistically.TrendAI report, framing section and unresolved questions
- Secondary accounts add claims the TrendAI report does not make. The Hacker News (2026-10-27) quotes TrendAI in English and says CyberStrikeAI was earlier used in a campaign against Fortinet FortiGate appliances; the CSA note (2026-10-28) attributes that link to Team Cymru and cites an English TrendAI version dated 2026-10-27; this record relies on the Japanese report. The Hacker News also says it is unknown whether the attacks reached other regions, while TrendAI names targets in Taiwan, Hong Kong and other countries. The CSA note says it was produced with AI assistance and not through CSA's official review.The Hacker News article; CSA research note, Background and references
Record updates
Correction: repaired a summary sentence that conflated directly supported attacker-side success evidence with unresolved damage, operator identity and model attribution.
Findings that cite this record
Key questions this bears on
- How are attackers using AI agents in real operations?To run large parts of intrusions under human direction, including espionage, credential theft and extortion; reported autonomy and success vary.