China-based group used AI to target more than 200 organizations, including Latin American foreign ministries
According to CrowdStrike, the VAPOR PANDA group used artificial intelligence agents to target more than 200 organizations in 11 days, including foreign ministries in Latin America. The attackers did not use new techniques. Instead, AI allowed them to run many intrusions at the same time, which changes defense priorities for companies in the region.
Photo: Shahadat Rahman on Unsplash
Over eleven days in October, a single person directing an artificial intelligence program carried out dozens of cyberattacks simultaneously. That is how cybersecurity company CrowdStrike describes the campaign by VAPOR PANDA, a China-based group active since 2021. In a report published on its blog on October 29, 2026, the firm explains that the group targeted more than 200 organizations in at least 14 countries. Among them were Latin American foreign ministries.
The group used ARTEX, an open-source framework for building AI “agents.” An agent is a program that doesn't just answer questions but also chains actions together on its own: it scans systems, tries passwords, runs commands and decides on the next step. According to CrowdStrike, the brain behind those agents was mainly DeepSeek, the language model developed in China. The firm did not specify which other commercial model was used. The activity began on October 17 from a single point of control.
What AI changed: speed
The most striking finding is not a new attack technique, because there wasn't one. CrowdStrike stresses that the methods were already known and that what was new was the speed and the ability to do many things at once. Adam Meyers of CrowdStrike told CyberScoop that a human attacker typically runs up to 40 commands per hour. With ARTEX, that figure could exceed 1,500. He also said it is the first time they have seen a state-backed actor use agents at this scale. The company recorded about 45 attacks of this kind in September and 126 in October.
The machine's workload was also enormous. In a single intrusion, the agents that planned the steps processed about 1.4 billion *tokens*, the fragments of text that an AI model reads.
How many victims were actually compromised is the least clear part. CrowdStrike's report confirms so-called post-exploitation activity (attacker activity once already inside the systems) at nine organizations. It also documents *password spraying* against at least 12 others. This technique involves trying common passwords across many accounts at once. CyberScoop, by contrast, refers to about 140 compromised targets. According to that outlet, CrowdStrike told it that nearly 70% showed compromise or signs of intrusion with “moderate confidence.” The company's blog does not mention that figure, and no source has explained how the two measurements relate to each other.
In addition to the Latin American foreign ministries, the reported targets include a presidential office and a prime minister's office in Europe, defense and intelligence agencies, 19 organizations in strategic sectors in the United States, and telecommunications companies in the United States, Belgium and the Philippines.
From assistant to operator
This case marks another step in an evolution that had already been observed. In 2024, Microsoft and OpenAI described state-backed groups using AI as a mere assistant, with capabilities that OpenAI called “limited, incremental.” In 2025, Anthropic reported that a group it linked to China, GTG-1002, had used its Claude Code tool almost autonomously against about 30 organizations, although outside experts questioned those claims. And in February 2026, Google reported that actors from China, North Korea, Iran and Russia were using its Gemini model in every phase of an attack.
VAPOR PANDA expands that scale and does so with a reusable tool. CrowdStrike says it had already seen ARTEX in the hands of financially motivated criminals and unidentified actors. In early October, the firm linked it to attacks on South Korean banks. It also distinguishes this campaign from activity that other companies in the industry track as Storm-0940. According to The Hacker News, ARTEX's creator closed the tool's source code after several groups began using it. The firm Zscaler had already warned about it.
The region was already familiar with this type of threat. According to Clarín and other media outlets, CrowdStrike attributed to a different attacker the use of AI agents with Claude Code to breach at least 15 Argentine healthcare entities between May and August 2026. It is a separate case, but it shows that this type of attack has already reached the region.
In Latin America, the report points to two foreign ministries with extensive intrusions and a third with a probable compromise. It also mentions target lists prepared against Central American foreign ministries. The countries have not been disclosed. Nor are there reports of private Latin American companies affected by this campaign, or public reactions from governments or regional incident response teams.
For companies, the lesson is less futuristic than it seems. If AI makes it possible to try doors at high speed, every lapse matters more. CrowdStrike recommends periodically rotating keys and credentials and using IMDSv2, a more secure configuration for cloud servers. It also advises properly validating single sign-on (SAML) systems, applying patches promptly and running endpoint detection and response (EDR) software on all devices. Open questions remain: which countries were affected, how much the AI did on its own versus the operator, whether data was stolen, and how DeepSeek and governments will respond. For now, all figures come from CrowdStrike and have not been verified by third parties.