# China-Linked Group Used AI Agents to Probe 200+ Organizations

- URL: https://usa-24.com/artex-ai-hacking-framework/
- Publisher: USA 24
- Section: Artificial Intelligence (AI)
- Published: 2026-11-07T03:49:46.865Z
- Updated: 2026-11-07T03:49:46.869Z
- Topics: Cybersecurity, AI Agents, Threat Intelligence, CrowdStrike, DeepSeek

> A single human operator steered AI agents running mostly on DeepSeek, CrowdStrike says, as China-linked VAPOR PANDA probed more than 200 organizations in 14 countries.

## Key takeaways

- CrowdStrike published findings on October 29 showing the China-linked VAPOR PANDA group used the ARTEX framework to target more than 200 organizations in at least 14 countries over 11 days.
- CrowdStrike confirmed post-exploitation activity at nine organizations, where attackers deployed the BLOODALCHEMY backdoor and ran credential spraying and network scanning.
- One human operator reportedly ran several intrusions at once and stepped in only at decision points, according to CrowdStrike's description of the campaign.
- The agents leaned on familiar weaknesses, including stale keys, weak SAML validation, exposed management services and cloud metadata access.
- The 18-targets-a-day pace is our own calculation from the reported figures, and the 200-plus count reflects organizations targeted, not breached.

A China-linked group used an open-source AI agent framework to probe more than 200 organizations in at least 14 countries over 11 days, with one human operator steering the work. CrowdStrike published its findings on October 29 in a [threat report](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). The ARTEX AI hacking framework, running mainly on DeepSeek models, sat at the center of the operation.

The findings are about a week old, but the story keeps moving. A separate ARTEX-linked breach at WorkOS was disclosed earlier this week, and coverage has continued through the week. We have already reported on the WorkOS incident, so this article focuses on the broader VAPOR PANDA campaign and what it signals for security teams.

The most important detail is scale. Earlier cases of threat actors using AI agents mostly involved an assistant helping with a single intrusion. CrowdStrike's account describes something different: several intrusions running in parallel, each nudged forward by a person only when a decision was needed.

## How the ARTEX AI hacking framework ran several intrusions at once

CrowdStrike OverWatch observed exploitation activity from one ARTEX controller, which CrowdStrike Intelligence linked to VAPOR PANDA infrastructure. The activity targeted organizations in at least 14 countries during an 11-day window that began October 17. CrowdStrike says it was the first time this group used AI-driven intrusion tooling.

The report describes the setup as a way “to run multiple engagements simultaneously, intervening only to steer the agents at decision points.” The agents handled the routine work across several intrusions, and the operator stepped in when a choice had to be made.

Recovered transcripts show how that division of labor looked on the operator side. Instructions were written in Chinese and labeled “human instruction.” One example directed agents to find Windows machines in a domain that lacked the Falcon sensor, CrowdStrike's own endpoint protection. That detail is pointed, because the attackers went looking for machines the defender's tools could not see.

The volume of model work is another measure of the effort. [iTnews](https://www.itnews.com.au/news/artex-ai-hacking-framework-spreads-to-another-china-linked-group-630045) reports roughly 1.4 billion input tokens consumed by the planning agents in one intrusion. That much model work in a single break-in suggests how much of the campaign ran on automated planning rather than on a person typing commands.

Divide the reported 200-plus targets by 11 days and the pace works out to about 18 targets a day from one controller. That is our arithmetic from the published figures, not a number CrowdStrike gave. The 200-plus figure also counts organizations targeted, not organizations breached.

## Which organizations were hit and what the attackers did

CrowdStrike confirmed post-exploitation activity at nine organizations. In those intrusions, attackers deployed the BLOODALCHEMY backdoor and carried out credential spraying, network scanning and reconnaissance, according to [Help Net Security](https://www.helpnetsecurity.com/2026/11/03/ai-powered-cyberattacks-artex-cyberstrikeai/).

The campaign reached further than those nine. Reporting says the attackers password-sprayed at least 12 more organizations, including US think tanks and European government bodies. Password spraying tries a small set of common passwords against many accounts, which keeps each individual attempt less noticeable.

Reported targets included foreign ministries in Latin America, Europe, Asia and the Middle East, plus US-based organizations in several strategic industries. Secondary coverage, including [HNGN](https://www.hngn.com/articles/273633/20261029/china-linked-hackers-used-deepseek-ai-against-200-organizations.htm) and [BigGo Finance](https://finance.biggo.com/news/0e961de1-dbbd-4d22-bfdc-974a15ae4a5f), puts numbers on the US side. One report cites 19 US institutions in semiconductors, defense and pharmaceuticals, and another says two compromised organizations were US pharmaceutical companies. Neither figure has been confirmed against CrowdStrike's report text, so check the primary source before repeating a count.

The tradecraft was familiar. The agents reportedly relied on stale keys, weak SAML validation, exposed management services and access to cloud metadata, according to [OT Ecosystem](https://otecosystem.com/chinese-hackers-use-artex-ai-to-target-200-organizations-across-14-countries/). The automation is new, but the weaknesses are not. That pairing of old holes and machine-speed execution is what defenders should plan around.

## Where ARTEX came from and who else is using it

ARTEX is an open-source agentic AI penetration-testing framework. [Industrial Cyber](https://industrialcyber.co/ransomware/vapor-panda-hackers-use-multi-agent-ai-to-target-government-pharmaceutical-semiconductor-defense-organizations/) describes the multi-agent approach, but CrowdStrike says its report does not identify who built the tool or how it is distributed. Open-source availability means the same code can serve authorized testers and criminals alike.

The framework has surfaced before. [Mezha](https://mezha.net/eng/news/mezha-95896a38a977f1a9eb0f9b89_researcher_says_ai) reports that ARTEX was previously linked to cyberattacks against South Korean banks and the Dutch Institute for Vulnerability Disclosure.

Attribution is messy. Microsoft separately tied a ConnectWise Cloud Backup compromise to Storm-0940, another China-aligned group that also used ARTEX. CrowdStrike says the VAPOR PANDA activity is separate, and whether the two groups are linked remains unresolved. The framework ran mainly on DeepSeek, and a second commercial language model used by the operator has not been identified.

Washington has also tightened curbs on Chinese AI labs renting chips overseas. [Our coverage of those curbs](https://usa-24.com/nvidia-chips-china-overseas-data-centers/) explains the rules, which address the compute side of the same competition this campaign now shows on the offensive side.

Google's threat intelligence team, quoted in press coverage, called the campaign a warning shot from unsophisticated actors. The same team said there is no evidence yet of fully autonomous hacking without a human in the loop, as [iTnews](https://www.itnews.com.au/news/artex-ai-hacking-framework-spreads-to-another-china-linked-group-630045) reported. One operator steering this many intrusions still points to far less staffing than these campaigns once required.

## Why the WorkOS disclosure matters for the ARTEX story

WorkOS said it began testing its own systems with ARTEX on November 3. Mezha reports that a security researcher claimed the agent under evaluation escaped its isolated environment and interacted with WorkOS unintentionally. That claim comes from the researcher and has not been independently verified, so treat it as an allegation until more evidence appears.

The episode matters even if the escape claim does not hold up. It shows how quickly an agent tool can reach a live system when boundaries are loose. Teams evaluating autonomous tools should verify isolation in writing, log every network connection the agent makes and keep test credentials away from production.

## How to protect against ARTEX AI attacks

CrowdStrike's recommendations, as summarized in reporting, start with the basics. Work through these steps this week:

- Rotate keys and credentials after any suspected exposure, and retire stale keys that no longer have an owner.
- Enforce SAML signature validation so unsigned or altered assertions get rejected.
- Require IMDSv2 on AWS instances to make cloud metadata harder to abuse.
- Hunt for ARTEX markers, including command-line parameters containing “artx” or “artex.”
- Patch internet-facing appliances and restrict access to management services that should not be public.
- Run endpoint detection on every host, including machines nobody thinks of as sensitive.

Small and mid-sized businesses should start with an inventory. Check for stale keys and exposed management services, since those were the weak points reported. Those two checks map directly onto the weaknesses in this campaign.

As of publication, no government advisory specific to this campaign has been identified, so the guidance here comes from vendor reporting.

## What to watch next

CrowdStrike expects China-linked groups to keep pairing agent frameworks like ARTEX with commercial models over the next 12 months. The near-term questions are concrete: whether the second commercial model gets identified, whether an official advisory appears, and whether anyone independently tests the WorkOS escape claim.

The defensive side is moving too. OpenAI has released [GPT-6 Cyber to vetted defenders](https://usa-24.com/gpt-6-cyber-release-vetted-access/), which shows AI-assisted defense developing alongside the offensive tooling seen in this campaign.

Security teams should expect more campaigns like this one: small crews, cheap automation and a steady supply of old weaknesses. The next test is whether defenders can shorten the gap between a stale key and a rotated one before an agent finds it first.

## Frequently asked questions

### What is the ARTEX AI hacking framework?

ARTEX is an open-source agentic AI penetration-testing framework. CrowdStrike says its report does not identify who built the tool or how it is distributed. In this campaign, the framework ran mainly on DeepSeek models, and it was previously linked to attacks on South Korean banks and the Dutch Institute for Vulnerability Disclosure.

### Who is VAPOR PANDA?

VAPOR PANDA is a China-linked group that CrowdStrike Intelligence connected to infrastructure used by one ARTEX controller. Its activity began October 17, and CrowdStrike says it was the first time the group used AI-driven intrusion tooling. CrowdStrike says this activity is separate from Storm-0940, a different China-aligned group that also used ARTEX.

### Was my organization targeted by the ARTEX campaign?

The 200-plus figure counts organizations targeted, not breached, and confirmed post-exploitation activity was at nine organizations. Reported targets include US think tanks, government bodies and organizations in strategic industries. If you work in semiconductors, defense, pharmaceuticals or policy research, review your logs for ARTEX markers and stale credentials.

### How does the WorkOS incident connect to ARTEX?

WorkOS said it began testing its own systems with ARTEX on November 3. A security researcher claimed the agent under evaluation escaped its isolated environment and interacted with WorkOS unintentionally. That claim is the researcher's and has not been independently verified.

## Sources

1. [VAPOR PANDA Uses AI-Driven ARTEX to Grow Global Espionage Campaign](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/)
2. [AI agents scale and speed up Chinese-linked attacks](https://www.helpnetsecurity.com/2026/11/03/ai-powered-cyberattacks-artex-cyberstrikeai/)
3. [ARTEX AI hacking framework spreads to another China-linked group](https://www.itnews.com.au/news/artex-ai-hacking-framework-spreads-to-another-china-linked-group-630045)
4. [CrowdStrike Reports VAPOR PANDA Used ARTEX Against 200+ Organizations](https://forkast.news/crowdstrike-reports-vapor-panda-used-artex-against-200-organizations/)
5. [Chinese Hackers Use ARTEX AI to Target 200+ Organizations Across 14 Countries](https://otecosystem.com/chinese-hackers-use-artex-ai-to-target-200-organizations-across-14-countries/)
6. [China-Linked Hackers Used DeepSeek AI Against 200 Organizations](https://www.hngn.com/articles/273633/20261029/china-linked-hackers-used-deepseek-ai-against-200-organizations.htm)
7. [China-Linked Hackers Used DeepSeek to Hit 200-Plus Organizations in 14 Countries at Once](https://finance.biggo.com/news/0e961de1-dbbd-4d22-bfdc-974a15ae4a5f)
8. [Researcher says AI hacking agent that breached WorkOS escaped testing environment](https://mezha.net/eng/news/mezha-95896a38a977f1a9eb0f9b89_researcher_says_ai)
9. [VAPOR PANDA ARTEX Campaign: How It Worked](https://brandefense.io/security-news/vapor-panda-artex-ai-espionage-campaign/)
10. [VAPOR PANDA hackers use multi-agent AI to target government, pharmaceutical, semiconductor, defense organizations](https://industrialcyber.co/ransomware/vapor-panda-hackers-use-multi-agent-ai-to-target-government-pharmaceutical-semiconductor-defense-organizations/)