A week ago I wrote about ARTEX showing up twice in 24 hours, once in Zscaler ThreatLabz research and once in Microsoft's ConnectWise investigation. (Usual disclosure: I work at Zscaler.) This week the story moved from "researchers found an AI hacking framework" to two things that are a lot more concrete.
Vapor Panda Ran ARTEX at Scale
On October 29, CrowdStrike published research on VAPOR PANDA, a China-nexus state-sponsored group (overlapping with Trend Micro's Earth Estries), using ARTEX to run intrusions in parallel. The numbers:
- Activity starting October 17, against 200+ organizations in at least 14 countries over 11 days.
- Confirmed post-exploitation at 9 organizations, with password spraying and credential scanning against at least 12 more.
- Targets included foreign ministries across four regions, U.S. pharma, semiconductor, defense, nuclear and think tank organizations, and telecoms in the U.S., Belgium and the Philippines.
How it worked: a main agent coordinates planner and worker agents that share a graph of findings and goals, with a human operator stepping in at decision points. The main model was DeepSeek, plus an unidentified commercial LLM. One intrusion's planning agents chewed through roughly 1.4 billion input tokens. CrowdStrike says this is the group's first observed use of AI-driven intrusion tooling.
What I'd underline is the techniques. They weren't magic. They were an unsigned SAML assertion chained with an XXE bug to reach cloud metadata, an unauthenticated orchestration service, LAPS passwords, and, at one government target, signing keys that hadn't been rotated in years. The operator even told the agents to find machines without an EDR sensor. The AI didn't invent new attacks. It found the old mistakes faster and in more places at once.
Chainalysis Says It Was Hit by "Offensive Agentic Techniques"
On October 30, Chainalysis, the blockchain analytics company that law enforcement and exchanges rely on, disclosed a security incident. According to BleepingComputer, the company told customers an external actor used "offensive agentic techniques" to access internal systems and tools, and that user IDs, names, email addresses and role labels for provisioned users were accessed. Chainalysis's public notice is shorter: it says it has found no evidence of access to investigation content, credentials or session tokens, is working with outside experts, and that this is unrelated to the Binance theft.
What we don't know: how the attacker got in, what was exploited, how long they had access, or what "agentic" meant in practice. I'm not going to guess. But "agentic" now appearing in a breach notice from a security-adjacent company is its own milestone.
Why This Matters
Agentic attackers change the math on breadth, not on cleverness. A small team can now work 200 targets at once and try every path in parallel, switching when one fails. That means the one weak spot you've been meaning to fix (the stale key, the exposed service, the server without a sensor) gets found. Not eventually. This month.
What To Do
- Rotate signing keys and service credentials on a schedule, and immediately after any suspected exposure. Years-old keys were the shortcut to admin here.
- Enforce SAML signature validation and IMDSv2 in cloud environments.
- Get EDR on everything, including the internal servers nobody thinks about. The attacker literally went looking for the gaps.
- Hunt for ARTEX markers CrowdStrike published, like
artxorartexin command lines andC:\ProgramData\artx5v\. - Shrink what's reachable. An agent can only test paths it can see. Apps that aren't exposed to the internet, and users who connect to apps instead of networks, leave a lot fewer paths to test.
- Chainalysis customers: turn on 2FA and be suspicious of password reset prompts and anyone "from Chainalysis" offering fund recovery or jobs.
Bottom Line
AI-run intrusions aren't a forecast anymore. They're in CrowdStrike's telemetry and in breach notices. The defense is unglamorous: fewer exposed paths, fresher keys, sensors everywhere. Same homework, just due sooner.
Sources: CrowdStrike on VAPOR PANDA and ARTEX, The Hacker News, BleepingComputer on Chainalysis, Chainalysis security incident notice, and Zscaler ThreatLabz ARTEX research