Threat researchCisent
AI agents, directed by a single person, broke into the foreign ministries of two Latin American countries
According to CrowdStrike, a China-linked espionage group had AI agents do the intrusion work at many organizations at once, while one person guided them at key decision points.

On October 29, 2026, CrowdStrike published what it saw of Vapor Panda, an espionage group it links to China: since October 17, its AI agents had been attacking organizations in at least 14 countries at the same time, and a single person was directing them.1 Among them were the foreign ministries of two Latin American countries, where the group carried out extensive intrusions, and probably a third.1
What happened
In 11 days, the campaign targeted more than 200 organizations, although CrowdStrike does not explain how it arrived at that figure. The targets included foreign ministries in Latin America, Europe, Asia and the Middle East; US semiconductor, defense, nuclear, pharmaceutical and financial companies; think tanks, European government bodies and telecommunications providers.1
CrowdStrike confirmed post-exploitation activity at nine organizations, where the group installed its backdoor, BLOODALCHEMY, and saw password spraying or credential scanning against at least twelve more. It also says the group compromised telecommunications providers in the United States, Belgium and the Philippines.1
How the agents worked
They used ARTEX, a framework in which a main agent coordinates other agents that plan and execute, over a shared record of facts, findings and goals. The agents moved from initial access to reconnaissance of each machine and lateral movement, tried several paths at once and switched tactics when one failed.1
The person stepped in mostly at decision points. In the conversations CrowdStrike recovered, the operator asked the agents to write web shells, extract the domain administrator passwords and find out which machines had security sensors. In a single intrusion, the planning agents processed about 1.4 billion input tokens.1
The main model the group used was one developed in China. It also used a second commercial, subscription-based model that CrowdStrike could not identify.1
How they got in
CrowdStrike describes some of the entry points:1
- at a US pharmaceutical company, an unsigned SAML assertion and an XXE flaw opened the cloud metadata service to them, and from there they reached an administrator account;
- at another pharmaceutical company, web shells and an unauthenticated orchestration service left them as root in a container environment;
- at a Greek government body, they forged an administrator token with keys from a 2022 directory export that were never rotated, and used VPN configurations and a client key leaked in May 2024.
Why it matters
For CrowdStrike it is a leap: from AI-assisted intrusions, one at a time, to several simultaneous ones directed by a single operator. It considers it highly likely that other China-linked groups will adopt frameworks like ARTEX over the next twelve months.1
ARTEX had appeared before. DIVD, the Dutch institute that warns organizations about their exposed flaws, said the tool used in the attack it suffered in September, which we wrote about on October 1, was ARTEX, and that it acted autonomously.2 Microsoft has also reported ARTEX activity by another group, Storm-0940, which CrowdStrike treats as separate from this campaign.1
What isn’t known yet
- Which foreign ministries they are, and what information the group took.
- How CrowdStrike arrived at more than 200 organizations, and whether the nine confirmed include the telecommunications providers.
- Which second model the group used.
Where Cisent comes in
Among the entry points CrowdStrike describes are password spraying and keys that were never rotated, and the agents changed course as soon as one path failed.1
Sentinel Defense detects, assesses and blocks attacks in real time alongside the EDR, the firewall and every connected source. A scan or a password campaign shows up while it’s happening, not months later. It acts on its own only where an action can be undone, such as blocking an attacking IP or an intrusive connection, and asks before anything with impact, such as isolating a machine.
Sources
CrowdStrike Counter Adversary Operations, “VAPOR PANDA Uses AI-Driven ARTEX to Scale a Global Espionage Campaign”, October 29, 2026. https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/↩︎
Bill Toulas, BleepingComputer, “Chinese hackers use ARTEX AI in automated attacks against 200 orgs”, November 2, 2026. https://www.bleepingcomputer.com/news/security/chinese-hackers-use-artex-ai-in-automated-attacks-against-200-orgs/↩︎