1.4 Billion Tokens Per Intrusion
CrowdStrike names Vapor Panda as the operator behind an ARTEX-driven campaign that reached 200+ organizations starting October 17.
On October 30 The Hacker News reported that CrowdStrike attributes a campaign running since October 17, 2026 to Vapor Panda, a China-linked state-sponsored group that overlaps Trend Micro's Earth Estries. The group used ARTEX — a Go-based AI penetration-testing framework the digest has been tracking since the July South Korean banks disclosure — against more than 200 organizations worldwide, including foreign ministries, U.S. semiconductor and defense firms, financial firms, and telecom providers. CrowdStrike reports post-exploitation confirmed at nine organizations, with password spraying and credential scanning observed against at least twelve more.
The specific numerator worth reading twice: planning agents inside the framework consumed roughly 1.4 billion input tokens in a single intrusion.
What the architecture actually is
Per the Hacker News piece, the ARTEX setup is a multi-agent system where a human operator oversees a team of planner and worker AI agents. The models in use: DeepSeek plus an unidentified subscription LLM. Agents pursue multiple attack paths in parallel and switch approaches when one fails, while the human operator steers them and works to bypass endpoint defenses.
In other words: the operator is not driving each command. The operator is driving the agent that is driving each command.
That difference shows up in the shape of the intrusion. The campaign's observed primitives, per CrowdStrike as summarized in the article:
- Cloud identity. An unsigned SAML assertion combined with an XXE flaw was chained to query the cloud Instance Metadata Service. The attacker got a role able to issue commands to thousands of instances, then escalated that to an enterprise administrator account.
- Containers. Web shells deployed, and an unauthenticated orchestration service was abused to gain root in a production container environment.
- Active Directory. LAPS passwords were read. Resource-based constrained delegation was used to impersonate a privileged account.
- Mail and VPN. Authentication keys unrotated since 2022 were used to forge an administrator token and read mailboxes. VPN configurations from sent mail and a VPN client key found in a May 2024 email leak were reused to reach the internal network.
- Post-exploitation implant. BLOODALCHEMY, a ShadowPad derivative.
The AD and mail steps are standard tradecraft. The specific thing to notice is the operating cadence: the agents are iterating these in parallel, with the human steering, over the course of a single intrusion and burning through roughly 1.4 billion input tokens of planner context to do it. That is not one operator typing one command at a time.
Four new Panda groups land in the same report
CrowdStrike also newly documented four more Panda groups alongside Vapor Panda: Hollow, Overcast, Titan, Veiled. The report mentions them without attributing them to the Vapor Panda campaign — the sibling-group documentation is the context, not the attribution. The relevant signal for defenders isn't the number of groups; it's that CrowdStrike is now publishing them at the per-group level in the same operating neighborhood.
The ARTEX developer, per earlier coverage already in the digest, reportedly took the tool closed-source after the South Korean banks disclosure. Closing the source doesn't retrieve the already-shipped copies. ARTEX is now a named-and-tracked framework whether the author wants it to be or not.
The point of the token-budget numerator
A per-intrusion token budget is a cost. It converts an attacker's work into a dollar figure. Depending on which subscription LLM is in the mix and which pricing tier it runs at, 1.4 billion input tokens is on the order of hundreds to low-thousands of dollars for the planner side of a single intrusion. Pair that with the economics in the October 21 CSA/ProjectDiscovery weight-poisoning note — an $8 QLoRA adapter cost to implant a working trigger-fire backdoor — and the operator-side budget to drive a sophisticated attributed intrusion end-to-end is no longer a mystery.
That's the shift. For the first time, we have a per-intrusion input-token numerator on a state-attributed agentic campaign — a budgeting primitive, not a vibe.
What the operator steered toward
Three of the technical choices in the campaign are worth lingering on because they tell you where the defender's attention needs to be:
- The 2022 authentication key that forged the admin token and read mailboxes. Unrotated identity material from four years ago was the entire pivot. The agentic cadence just accelerated the search for which forgotten key still works.
- The May 2024 VPN client key reused from a prior email leak. The intrusion's own data flow became an attack surface for the intrusion.
- The unsigned SAML + XXE → IMDS chain. A classical cloud identity primitive, driven at an unusual tempo. The AI didn't invent the chain. It just stopped the chain from waiting on a human.
Each of these would be a credible defender signal on its own. The specific defender problem is that the agents are running them in parallel, backtracking when they don't work, and burning input context at a pace that no human operator would.
What to actually do
- Rotate identity material you have forgotten about. The 2022 authentication key is the operator-side canary for this entire class of campaign. If you have signing keys, service account tokens, or integration credentials you haven't rotated since Trump's first term, this is what agents are searching for.
- Audit SAML signing policies for unsigned-assertion acceptance. The IMDS pivot in this report rode an unsigned SAML + XXE primitive that is a 2015-era defender problem being driven at 2026 pace.
- Treat parallel-attack-path probing against your identity surface as its own observable. The specific signature of an ARTEX-shaped operator is that failed attempts don't slow it down; they just fork off the next path. Rate-limit assumptions built for human operators are the wrong rate limit.
The takeaway
CrowdStrike put a named state-backed operator behind an in-the-wild agentic campaign that reached 200+ organizations, confirmed post-exploitation at nine of them, and burned roughly 1.4 billion input tokens of planner context per intrusion. The exploit primitives are 2015-era. The cadence is not.
The 2022 unrotated authentication key is the specific thing someone in your environment has forgotten about. The agents are searching for it, in parallel, right now.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.