CrowdStrike Says China-Linked Hackers Ran AI Agents Against 200-Plus Organizations

CrowdStrike reports that the China-nexus group VAPOR PANDA used a multi-agent framework called ARTEX, running mainly on DeepSeek models, to target more than 200 organizations in at least 14 countries over 11 days, with nine confirmed post-exploitation cases.

Security and governanceCrowdStrikeDeepSeekAI Agents

CrowdStrike says a China-linked espionage group ran one of the largest AI-driven intrusion campaigns documented so far, using a multi-agent framework to pursue more than 200 organizations across at least 14 countries in an 11-day window. The campaign, described in an October 29, 2026 report from CrowdStrike Counter Adversary Operations, was attributed to a group the firm calls VAPOR PANDA, a targeted-intrusion actor it has tracked since 2021. The framework at the center of the campaign is called ARTEX, and CrowdStrike says it ran primarily on models from the Chinese developer DeepSeek, alongside a second, unidentified commercial model.

The scale is the headline, but the confirmed damage is narrower, and the distinction matters. CrowdStrike says more than 200 organizations were targeted between October 17 and October 28, yet it confirmed post-exploitation activity at nine organizations. At least 12 more organizations faced credential spraying or scanning without confirmed compromise. That gap between targeting and breach is normal in espionage reporting, and it is worth stating plainly: ARTEX gave the operators reach, not automatic success everywhere it pointed.

ARTEX is what makes the campaign different from an ordinary intrusion set. CrowdStrike describes a controller that coordinates autonomous planner and worker agents sharing a single task graph, with a human operator steering at decision points rather than typing every command. Transcripts observed by the firm show Chinese-language commands from human operators instructing the agents to extract administrator credentials and deploy web shells. The planner agents, which break a goal into steps and assign them, consumed roughly 1.4 billion input tokens during a single intrusion, according to CrowdStrike. That figure is a measure of how much reasoning the campaign threw at one target, and it explains why defenders found the activity so hard to bound: the agents could keep generating, testing, and revising approaches without waiting for a human to catch up.

At organizations where the intruders got in, CrowdStrike reports a familiar espionage playbook executed at unfamiliar speed. The attackers installed a backdoor the firm calls BLOODALCHEMY, scanned internal networks, and directed credential-spraying efforts against further targets. One engagement alone targeted 19 United States-based organizations across pharmaceutical, semiconductor, defense, nuclear, and financial sectors. Other targets included European executive offices, a Middle Eastern interior ministry, and Latin American foreign ministries. The target list reads like a classic state-espionage collection plan. What is new is the machinery used to work through it.

The model choice is the part of the report that will draw the most scrutiny. CrowdStrike identifies DeepSeek as the primary model behind ARTEX, with an unidentified subscription-based commercial model as the secondary. The firm does not claim that DeepSeek built, directed, or approved the campaign, and nothing in the reporting supports that reading. The models were tools in the operators hands, the same way stolen credentials and commercial infrastructure are tools. But the detail still matters for policy, because it shows that frontier-capable models available to anyone can be wired into an autonomous intrusion pipeline by a determined actor, without any special access to the model provider.

CrowdStrike assessment is blunt about what comes next. The firm expects China-nexus adversaries will highly likely continue pairing commercial large language models with agentic frameworks, treating AI agents as a force multiplier for intrusion operations rather than an experiment. That judgment, if right, changes the defender arithmetic. A campaign that once required a room of operators can now be run by a small team supervising agents, and the bottleneck shifts from human typing speed to token budgets, infrastructure, and detection of agent behavior itself.

The report also sets a limit on what can be concluded from it. CrowdStrike attribution rests on its own telemetry, transcripts, and infrastructure analysis, and the second model in the campaign remains unidentified. Independent researchers have not yet published a matching technical analysis, and the targeted organizations have not, for the most part, spoken publicly. Those caveats do not shrink the campaign, but they do mark the boundary between what one well-resourced security firm observed and what the wider community has verified.

For defenders, the practical signal is about detection rather than any single indicator. Blocking one model, one backdoor name, or one set of infrastructure misses the point of ARTEX, which is that the planning layer can adapt. The campaign shows an espionage group treating AI agents the way earlier groups treated botnets: as a way to scale a small team across many targets at once. The organizations that were only sprayed or scanned, and never breached, are part of that story too. They show that the same machinery can probe hundreds of doors in days, and that the difference between a near miss and a breach may come down to defenses that were already in place before the agents arrived.