Chinese hackers use ARTEX AI in automated attacks against 200 orgs

What it is

ARTEX is a multi-agent AI penetration-testing suite built by Chinese developers. It was first released as open source, and its creators later made it proprietary after it was used in real attacks, including a campaign against South Korean financial institutions. CrowdStrike tracks the actor behind the October 2026 campaign as Vapor Panda. The campaign began on 17 October and targeted government ministries and organizations in telecommunications, pharmaceuticals, semiconductors, defense, nuclear and finance.

How it works

Vapor Panda used DeepSeek and a second, unidentified commercial model to run planning agents and worker agents across several intrusions at once. Operators wrote instructions in Chinese and stepped in at decision points. The agents did reconnaissance, exploited vulnerabilities, deployed web shells, collected credentials and switched to other attack paths when one failed. In one intrusion CrowdStrike believes the agents used more than 1.4 billion input tokens.

In the attack on a Greek government entity, the agents forged an administrator token and read mailboxes. They used authentication keys found in a 2022 mail-server directory export.

CrowdStrike says the actor also deployed BLOODALCHEMY, a backdoor active since 2023, against government entities in Latin America and the Middle East.

Preconditions / Affected

Confirmed victims include telecommunications providers in the United States and Belgium, government bodies in Greece, the Middle East and Latin America, and pharmaceutical firms in the United States. At least nine organizations were compromised. Password spraying and credential scanning hit at least 12 other high-value targets.

Operators told the agents to skip Windows hosts protected by CrowdStrike Falcon and to focus on unprotected hosts. Hosts without endpoint protection were therefore the intended targets.

DIVD case

The Dutch Institute for Vulnerability Disclosure says ARTEX acted autonomously in its September 2026 breach. The agent went from initial access to root in 12 minutes by chaining a Zammad zero-day with a second zero-day for local privilege escalation. It ran tasks in parallel: reading a compromised mailbox, scanning the internal network, extracting login tokens and attempting password resets. DIVD logged more than 450,000 requests and 2,500 scripts, and called the activity "super loud and messy."

Caveats

The page names no CVE ids for the two zero-days and gives no patch or mitigation guidance. It does not identify the second commercial model. The 1.4 billion token figure is a CrowdStrike belief, not a confirmed count. The DIVD details come from its LinkedIn post, as relayed by the article.