Skip to main content
All incident reports
Incident report

China-linked Vapor Panda used ARTEX AI framework against 200-plus organisations

CrowdStrike says a China-linked group it calls Vapor Panda used an AI-powered penetration-testing framework, ARTEX, against more than 200 organisations in 14 countries from 17 October 2026. Post-exploitation activity was confirmed at nine of them, and a Dutch disclosure body says the same tool was used against it in September.

· First published

The short version

  • The campaign. CrowdStrike attributes a wave of intrusions starting 17 October 2026 to a China-based actor it tracks as Vapor Panda, which used an AI-driven framework called ARTEX against more than 200 organisations in 14 countries. Targets included foreign and interior ministries, telecoms providers and defence, semiconductor, pharmaceutical, nuclear and financial organisations.
  • Confirmed intrusions. CrowdStrike confirmed post-exploitation activity at nine organisations, where the actor deployed BLOODALCHEMY, a ShadowPad derivative, then sprayed credentials and scanned networks. At least 12 further organisations, including US think tanks and European government bodies, were hit with password spraying and credential scanning.
  • How the tooling works. ARTEX is a multi-agent AI penetration-testing framework that coordinates planner and worker agents using DeepSeek and a second, unnamed commercial large language model (LLM). A human operator issues instructions in Chinese and intervenes at decision points; in one intrusion the agents consumed about 1.4 billion input tokens.
  • A separate ARTEX case. The Dutch Institute for Vulnerability Disclosure (DIVD) says an autonomous ARTEX agent went from initial access to root on its network in 12 minutes in September 2026, chaining a Zammad zero-day with a second privilege-escalation flaw. No source links that intrusion to Vapor Panda.

Timeline

DateEvent
February 2021Vapor Panda is first identified by CrowdStrike.
September 2026DIVD is attacked; it later attributes the intrusion to an autonomous ARTEX agent.
17 October 2026Vapor Panda begins using ARTEX against organisations worldwide.
30 October 2026CrowdStrike's analysis of the campaign is reported.
2 November 2026Further reporting adds DIVD's account of the attack on its own network.

What CrowdStrike found

CrowdStrike attributes the activity to Vapor Panda, a China-based group it first identified in February 2021 and which it links to the cluster Trend Micro calls Earth Estries. From 17 October 2026 the group used ARTEX, a Go-based framework that runs multiple agents against several targets at once, with a human operator steering them to write web shells, extract domain administrator passwords and map endpoint sensor coverage.

The agents chained known techniques rather than a single new flaw. Reported paths include combining an unsigned Security Assertion Markup Language (SAML) assertion with an XML external entity flaw to query the cloud Instance Metadata Service, abusing an unauthenticated orchestration service for root in a production container, reading LAPS passwords, and resource-based constrained delegation to impersonate a privileged account. In the Greek government intrusion, agents used authentication keys from a 2022 mail-server directory export to forge an administrator token and read mailboxes; VPN configurations recovered from sent mail and a VPN client key originating from a purported May 2024 email leak were then used to reach the internal network.

CrowdStrike says the operators instructed the agents to avoid Windows machines running its Falcon endpoint sensor and to concentrate on unprotected hosts, and that they switched attack paths when one failed. The firm's analysts reported about 1.4 billion input tokens consumed in a single intrusion.

Who is affected

CrowdStrike counts more than 200 organisations across 14 countries. The target list includes foreign ministries in Latin America, Europe, Asia and the Middle East, a Middle East interior ministry, and a European country's foreign ministry, presidency and prime minister's office alongside defence, justice, finance and intelligence bodies. It also names 19 US organisations in the pharmaceutical, semiconductor, defence, nuclear and financial sectors, and telecommunications providers in the US, Belgium and the Philippines.

BleepingComputer reports confirmed victims as telecoms providers in the US and Belgium, government organisations in Greece, the Middle East and Latin America, and US pharmaceutical firms. BLOODALCHEMY, described as a backdoor active since 2023, was deployed against government entities in Latin America and the Middle East.

DIVD and the spread of ARTEX

DIVD has published its own account of a September 2026 intrusion into its network, which it attributes to an autonomous ARTEX agent. It says the agent reached root in 12 minutes by chaining a Zammad zero-day with a second zero-day for local privilege escalation, and that it read a compromised mailbox, scanned the internal network, extracted login tokens and attempted password resets while generating more than 450,000 requests and 2,500 scripts. DIVD described the activity as "super loud and messy".

ARTEX began as an open-source project and was made proprietary by its developer after it was used in real-world attacks, including one against South Korean financial institutions, according to BleepingComputer. Zscaler's ThreatLabz, quoted by The Hacker News, said the framework lowers the technical barrier to complex intrusions.

The same CrowdStrike research names four previously undocumented China-linked clusters, Hollow Panda, Overcast Panda, Titan Panda and Veiled Panda, linked to intelligence gathering and intellectual property theft, including physical-access intrusions using a FlowCloud backdoor and attacks on legal billing software. No source connects those clusters to the ARTEX campaign.

What is not known yet

  • Which commercial large language model Vapor Panda used alongside DeepSeek; both sources leave it unidentified.
  • Whether data was taken from the nine organisations with confirmed post-exploitation activity, and how much; the sources describe mailbox access, web shells and a backdoor but give no exfiltration figures.
  • The complete victim list; the sources give sectors, regions and a few named countries only.
  • Whether DIVD's September intrusion was carried out by Vapor Panda or by another operator using ARTEX; no source links the two.
  • How widely ARTEX is now used by other actors, and whether other campaigns have been attributed to it beyond those named.

What to do

If you run internet-facing services

  1. Inventory externally reachable applications and confirm current patching for helpdesk and ticketing software, including Zammad, which DIVD names as one of two zero-days chained against it.
  2. Test how your applications handle SAML assertions and XML input, and reject unsigned assertions and external entity processing.
  3. Look for unauthorised web shells in web-accessible directories and review calls to orchestration services that accept unauthenticated requests.

If you run cloud, identity or mail infrastructure

  1. Rotate authentication keys and service secrets, and treat any key or directory export dating from 2022 or earlier as compromised; hunt for forged administrator tokens and unusual mailbox reads.
  2. Delete VPN configuration files and client keys sitting in mailboxes and sent mail, then rotate the corresponding VPN credentials.
  3. Audit reads of LAPS passwords and any resource-based constrained delegation set up to impersonate privileged accounts.

For everyone else

  1. Close endpoint coverage gaps and check that detection is working on every Windows host, since the agents were directed away from protected machines towards unprotected ones.
  2. Alert on password spraying and credential scanning against high-value accounts, which the sources describe as preceding or accompanying intrusions.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.