CrowdStrike: AI targeted 200-plus organizations across 14 countries

By Sophie Lindqvist

CrowdStrike said an AI-driven campaign targeted more than 200 organizations across 14 countries during an 11-day period. The firm attributed the operation to VAPOR PANDA, a hacking group believed to be linked to China.

The campaign began on the 17th of the previous month, according to a threat analysis CrowdStrike released on the 4th. The report supplied neither a calendar month nor a year.

Among the targets were 19 U.S. institutions spanning pharmaceutical, semiconductor, defense, nuclear and financial sectors. CrowdStrike did not name those institutions or specify which were compromised.

Which tools drove the campaign?

The agents used DeepSeek, a Chinese large language model, and ARTEX, an AI-based penetration tool. CrowdStrike's report said AI agents, rather than humans, carried out much of the attack.

Across the operation, agents divided their work among vulnerability scanning, system intrusion, internal reconnaissance and privilege theft, running those tasks in parallel and stage by stage. Attackers issued commands directly in Chinese. Humans intervened only at key decision points. Recovered logs instructed agents to build web shells, extract administrator passwords and hunt for machines without security solutions.

ARTEX began as open-source penetration-testing software designed to find vulnerabilities. Attackers repurposed it as a hacking tool. It had appeared in earlier attacks against South Korea's financial sector. During one intrusion, the agent that formulated the attack plan processed roughly 1.4 billion input tokens. That figure measures model input, not successful commands or compromised systems.

How broadly did the campaign spread?

The target set included U.S. semiconductor and defense contractors, pharmaceutical companies, nuclear institutions and financial organizations. It also included foreign ministries and presidential offices in several European countries.

Foreign ministries and agencies across Latin America, Europe, Asia and the Middle East were targeted in parallel, giving the operation broad geopolitical reach without establishing any link to one industrial sector. In one unnamed European country, attackers targeted:

• The foreign ministry
• The presidential office
• The prime minister's office
• Defense, justice, finance and intelligence agencies

The report did not provide country counts for most regions. That omission prevents a reliable ranking of the most affected jurisdictions.

Which compromises did CrowdStrike confirm?

Across at least nine organizations, attackers gained actual access, installed the BLOODALCHEMY malicious backdoor and attempted further intrusion, according to CrowdStrike's confirmed findings in the reported campaign.

At least 12 organizations showed repeated password attempts and credential discovery. Telecom carriers in the United States, Belgium and the Philippines were also compromised, although the report did not name them or give a count.

Those confirmed findings do not turn every target into a breach. The 19 U.S. institutions were targets; the source did not assign confirmed access to semiconductor sites.

What did the documented intrusions show?

• U.S. pharmaceutical company: An agent chained multiple vulnerabilities, reached a cloud system and extracted administrator privileges.
• Second target: An agent installed a web shell and obtained root-level administrator access inside a container environment.
• Greek government agency: Attackers used previously leaked credentials, forged administrator authentication and reached email and internal networks.

The three documented cases involved different entry paths—vulnerability chaining, web-shell deployment and credential abuse—rather than one repeatable method across every target. Each case confirms access, but the source gives no affected-data volume or count of compromised systems.

What does this mean for semiconductor companies?

The 19-institution U.S. figure spans five sectors. It is not a count of semiconductor breaches. CrowdStrike named no chip company and reported no fabrication outage, production loss or supply interruption. It also gave no evidence of wafer shipments, pricing or customer deliveries changing.

For chip suppliers, the report documents attempted intrusion against semiconductor-related institutions, not a quantified commercial loss or a demonstrated material change in fabrication capacity. Any claim that the campaign tightened chip supply or raised prices would exceed the evidence. The report's confirmed impact is limited to security exposure at unnamed institutions.

Can a small team sustain this scale?

Security experts assessed that AI-based cyberattacks had evolved beyond penetrating individual systems into campaigns capable of targeting multiple countries, institutions and sectors at the same time. They concluded that even a group with only a handful of members could reach far more targets than before by using AI agents.

That assessment remains an expert judgment, not measured staffing data for VAPOR PANDA. The source did not state how many operators participated or how many systems were compromised in total. DeepSeek and ARTEX lowered manual work by automating scans, intrusion steps and reconnaissance. Humans still intervened at decision points. Automation did not remove operator oversight.

What changes in the next 12 months?

CrowdStrike projected that China-linked groups are "highly likely to increasingly combine AI agent frameworks such as ARTEX with commercial AI models." The projection covers the next 12 months. It is a threat forecast, not a measured result.

The firm's analysis says AI lowers the entry barrier, raises automation and turns defender detection capabilities and response speed into critical variables for the next contest. The practical question is whether defenders can identify and contain parallel agent activity before credentials and privileged access spread.

Over the next 12 months, CrowdStrike expects more ARTEX-style combinations with commercial AI models, with detection speed and response time deciding whether attackers retain that scale advantage.