# 1.4 billion tokens for a single intrusion: VAPOR PANDA used ARTEX against 200+ organizations in 11 days

## Summary

On **October 29, 2026**, CrowdStrike published a cyberespionage campaign by **VAPOR PANDA** — a China-nexus actor active since 2021 — against **more than 200 organizations in at least 14 countries over 11 days**, starting **October 17, 2026**.

What is new is not the malware. It is that most of the attack work was **carried out by AI agents**, through an agentic penetration testing framework called **ARTEX**. The agents ran several attack paths in parallel, switched approaches when one failed, and handled exploitation, credential access and lateral movement across web apps, cloud, containers, Active Directory, mail and VPN infrastructure.

This is also the first time CrowdStrike has observed VAPOR PANDA using AI-driven intrusion tooling, and it came with **broadened targeting** into pharmaceutical, semiconductor, nuclear and defense sectors, plus foreign ministries in new regions.

One figure conveys the scale: **the planning agents alone consumed roughly 1.4 billion input tokens in a SINGLE intrusion**.

**Priority actions:**

- **Hunt for ARTEX markers now** — command-line parameters containing `artx` or `artex`, cloud roles prefixed `artx-`, `art18-` comments in administration commands, and the directory `C:\ProgramData\artx5v\`. These are framework markers, so they catch other groups using ARTEX too.
- **Audit every key and credential never rotated after an exposure**, including in backups and directory exports. Two of this campaign's five entry paths were old keys from 2022 and 2024 that still worked.

## What ARTEX is

ARTEX is an **open-source, model-agnostic agentic penetration testing framework** — it is orchestration, not a model.

The architecture as CrowdStrike describes it: **a main agent coordinates subordinate planner and worker agents** over a shared graph of facts, findings and goals. Each engagement, which ARTEX calls a "task", moves from initial access through host reconnaissance to lateral movement.

Per SlashID's analysis, the framework has a Go backend and a Next.js console, running a planner/worker loop over **two linked graphs**: an **asset graph** (domains, IPs, services, applications, endpoints) and an **exploration graph** (goals, actions, observations, findings). The planner creates work items, concurrent workers claim and execute them, and results feed the next planning round. SlashID also reports that the **public repository went private on October 8, 2026**, though forks circulate, and that ARTEX has been observed running on Claude, GPT, DeepSeek, GLM and Grok.

The deployment VAPOR PANDA used runs on **DeepSeek**, plus **a second subscription-based commercial LLM that CrowdStrike could not confirm the provider of**.

## About that 1.4 billion tokens

In a single intrusion, **the planning agents alone consumed roughly 1.4 billion input tokens**. CrowdStrike notes this covers planning only, so total consumption was considerably higher.

The figure is worth pausing on, because it describes the nature of the threat more precisely than any technical description: this is not one person typing commands and occasionally asking an AI. It is a system reading, reasoning and experimenting continuously at machine speed throughout the intrusion, with humans appearing only at the forks.

It also explains why 200 organizations in 11 days is achievable.

## What the AI did, and what the human did

| Handled by the agents | Directed by the operator |
| --- | --- |
| Running **multiple parallel intrusions**, each pursuing several attack paths at once | Issuing instructions **in Chinese**, labeled as human instructions and **timestamped in China Standard Time** |
| Switching to alternative approaches when a path failed | Intervening at decision points: writing web shells |
| Exploitation, credential access and lateral movement across web apps, cloud, containers, Active Directory, mail and VPN | Extracting domain admin passwords |
| **Endpoint sensor inventory** — identifying hosts without Falcon | Checking security sensor coverage |

The last item on the left is the one that matters most. **Inventorying which hosts lack EDR protection is now an automated reconnaissance step**, running at the same speed as everything else. The gaps in your EDR coverage are no longer something an attacker stumbles into — they are something the agents go looking for, from the start.

A small detail that points at the origin: CrowdStrike noted the operator using **PRC red-team slang for Falcon — "猎鹰"** (hunting falcon).

## Five ways in

CrowdStrike describes five distinct initial access vectors in this campaign. None of them is a zero-day.

**1\. Unsigned SAML chained with XXE.** An unsigned SAML assertion was accepted, then chained with an XXE flaw to reach the **cloud metadata service (IMDS)**. That yielded a role capable of **issuing commands to thousands of instances**, and from there an enterprise admin account.

**2\. An unauthenticated orchestration service.** Abused in a production container environment to gain root, after web shells had been deployed.

**3\. Stale keys never rotated.** A **2022 directory export from a Greek government mail server** contained authentication keys likely never rotated. They were used to **forge an admin token and read mailboxes**.

**4\. VPN from old email.** VPN configurations recovered from a user's sent mail, plus a **VPN client key from a May 2024 email leak**, used to reach the internal network.

**5\. Password spraying** against government, think tank and corporate accounts.

Vectors three and four deserve separate attention: both are **data leaked years earlier that still worked in 2026**. The agents did not have to break anything — they only had to read back what was already lying around.

## BLOODALCHEMY and the toolkit

**BLOODALCHEMY** is the primary backdoor, first publicly documented in October 2023. The load chain is classic sideloading:

```plaintext
svchost.exe                      # actually a renamed legitimate GameHook binary
   -> sideloads graphics-hook-filter64.dll    # malicious DLL
      -> decrypts graphics-hook_config.bin    # payload
         -> BLOODALCHEMY
```

Copies were staged in **open directories exposed to the internet**. The implant carries the internal path `D:\ProjectX\Merman_New\`.

The rest of the toolkit:

```plaintext
Web shells:       JSP and PHP
Credentials:      LAPS password reads
Escalation:       resource-based constrained delegation (RBCD) abuse for impersonation
                  forged admin tokens
Tunneling/proxy:  Chisel, SoftEther, Tor, SOCKS proxies, a compromised relay
```

Note that **nothing on this list is new**. All of it is known technique and known tooling. The only new thing is the speed, and the number of attack paths running at once.

## The markers ARTEX leaves behind

This is the immediately usable part, and the reason to read this report even if your organization does not fit VAPOR PANDA's targeting profile:

```plaintext
Command-line parameters containing:   "artx" or "artex"
Cloud roles prefixed:                 artx-
Comments in administration commands:  art18-
Directory:                            C:\ProgramData\artx5v\
```

Because ARTEX is a shared framework rather than one group's private tool, these markers **catch other groups using it as well** — something rarely true of conventional IOCs.

## Who was targeted

CrowdStrike says "more than 200 organizations" but itemizes only around 21 by name or type. Grouped by the level of certainty they state:

**Confirmed post-exploitation — 9 organizations.** The actor gained access, deployed BLOODALCHEMY, then conducted credential spraying, network scanning and reconnaissance.

**Password spraying or credential scanning — at least 12 further organizations**, including U.S. think tanks and European government entities.

**Foreign ministries:** two Latin American countries extensively intruded, a third Latin American country likely compromised, staged targets in Central America and Asia, plus a Greek government entity.

**Interior ministry:** a Middle East-based interior ministry was compromised.

**One European country targeted comprehensively:** in a single engagement, the agents simultaneously targeted that country's **foreign ministry, presidency, prime minister's office, and its defense, justice, finance and intelligence entities**.

**United States:** one engagement targeted **19 U.S. organizations** across pharmaceutical, semiconductor, defense, nuclear and financial sectors. Two U.S. pharmaceutical organizations were described specifically.

**Telecommunications:** providers in the **U.S., Belgium and the Philippines** were compromised.

**Global pharma:** one organization was targeted specifically for sensor-coverage evasion.

## Assessment

**Automation changes the economics of espionage, not just its speed.** More than 200 organizations in 14 countries over 11 days is a machine's number, not a person's. Target selection used to be bounded by the hours an operator had. Once that bound disappears, the selection logic shifts from "who is most valuable" to "try everything, keep whatever opens". The defensive consequence: **"we are not important enough to be targeted" has stopped being a valid self-assessment**.

**EDR coverage inventory is now an automated reconnaissance step.** This is the most concerning detail in the whole report. The agents actively look for hosts without a sensor, and they do it at machine speed. Every exception you ever made for performance reasons, for a legacy application that would not tolerate an agent, for an "unimportant" server — all of it is now being searched for systematically.

**Keys from 2022 and 2024 still opened doors in 2026.** Two of five entry paths were long-leaked credentials that were never rotated. This lesson recurs with alarming frequency in recent reports, and it is not a technical problem but a process one: **detecting a leak and finishing the rotation are two different tasks, and the second one often goes undone**.

**ARTEX is a shared tool, so hunting has to change accordingly.** In the last eight days of October alone, two reports covered two different groups using the same framework: on October 21 Microsoft published **Storm-0940** using ARTEX in the **ConnectWise** supply chain compromise, reaching more than 2,250 Microsoft 365 tenants at 2,900 sign-in attempts per minute; on October 29 came this report. **CrowdStrike states plainly that the two campaigns are not the same.** Microsoft Threat Intelligence adds that multiple actors from various regions are using ARTEX. In that context, IOCs tied to one group lose value quickly, while **framework markers catch all of them** — which is why the four `artx` indicators above are worth more than the three C2 domains.

**Nothing in the toolkit is new, and that is the point.** BLOODALCHEMY dates to 2023, DLL sideloading is old technique, Chisel and SoftEther are commodity tools, password spraying is textbook. The threat comes not from new technique but from **old technique executed simultaneously, tirelessly, across hundreds of targets**.

**CrowdStrike expects this to continue.** They assess that China-nexus adversaries will "highly likely" keep adopting agentic frameworks like ARTEX over the next 12 months. For an open-source, model-agnostic framework whose forks are already circulating, that is hard to argue with.

## Relevance for Vietnam

- **Vietnam does not appear in the victim list CrowdStrike published.** But **the Philippines does**, among the telecommunications victims, and the overall targeting profile — foreign ministries, telecoms, energy, defense — matches the kind of organization common across Southeast Asia.
- **Speed is the core problem for Vietnamese SOCs.** An attack chain running at agent speed does not wait for the next shift. If your process includes a "we'll look at it in the morning" step, that step needs reassessing in this context.
- **Hosts without EDR are actively hunted.** Many Vietnamese organizations still run legacy servers, old application hosts, or appliances that cannot take an agent. Write that list down — because the attacker's agents are compiling the same list.
- **Unrotated old keys are a very common exposure in Vietnam**, especially in systems that have passed through several administrators and several contractor handovers. Directory exports, config files sitting in email, VPN keys in handover documentation — all of it was successfully exploited in this campaign.
- **Password spraying at this scale produces a clear signal.** If your organization lacks an alert for the "many accounts, few passwords, short window" pattern, this is a low-cost, high-value detection to add.
- **Do not just block the three C2 domains.** They will change. The four `artx` markers are what hold their value, and they catch other groups using ARTEX too.

## Recommendations

- **Hunt for the ARTEX markers across the environment**: command-line parameters containing `artx`/`artex`, cloud roles prefixed `artx-`, `art18-` comments in administration commands, and the `C:\ProgramData\artx5v\` directory. Make these standing hunting rules, not a one-off sweep.
- **Rotate signing keys, preauthentication keys and service credentials after any suspected exposure** — including keys sitting in backups and directory exports. And treat "noted for rotation" as not a terminal state.
- **Enforce SAML signature validation.** The campaign's first entry path was an unsigned assertion being accepted.
- **Require IMDSv2** and **restrict the command permissions of cloud roles**. A role that can issue commands to thousands of instances is a permissions design flaw, not an incident.
- **Patch internet-facing appliances**, especially long-lived Linux servers that rarely get touched.
- **Deploy EDR across the whole environment**, including servers with no internet exposure. And **maintain a list of hosts that cannot take an agent**, with compensating controls for each.
- **Monitor for large-scale password spraying patterns**: many accounts, few passwords, concentrated in a short window, often from one IP range or through proxies.
- **Watch for commodity tunneling tools** — Chisel, SoftEther, Tor connections, SOCKS proxies — as anomalies on production servers.

## MITRE ATT&CK

> Mapping exactly as published by CrowdStrike.

| Tactic | Technique | Note |
| --- | --- | --- |
| Resource Development | **T1588.007** — Obtain Capabilities: Artificial Intelligence | Use of ARTEX and commercial LLMs |
| Initial Access | T1190 — Exploit Public-Facing Application | Unsigned SAML, XXE |
| Initial Access | T1078 — Valid Accounts | Stale keys and credentials |
| Initial Access | T1133 — External Remote Services | VPN |
| Execution | T1651 — Cloud Administration Command | Role commanding thousands of instances |
| Persistence | T1505.003 — Server Software Component: Web Shell | JSP and PHP |
| Privilege Escalation | T1134.001 — Access Token Manipulation: Token Impersonation/Theft | Via RBCD |
| Defense Evasion | T1574.001 — Hijack Execution Flow: DLL | BLOODALCHEMY load chain |
| Defense Evasion | T1036.005 — Masquerading: Match Legitimate Name or Location | GameHook renamed svchost.exe |
| Credential Access | T1552.005 — Cloud Instance Metadata API | IMDS |
| Credential Access | T1555 — Credentials from Password Stores | LAPS passwords |
| Credential Access | T1606.002 — Forge Web Credentials: SAML Tokens |  |
| Credential Access | T1606 — Forge Web Credentials | Forged admin tokens |
| Credential Access | T1110.003 — Brute Force: Password Spraying |  |
| Discovery | T1518.001 — Software Discovery: Security Software Discovery | Endpoint sensor inventory |
| Collection | T1114.002 — Email Collection: Remote Email Collection |  |
| Command and Control | T1090.003 — Proxy: Multi-hop Proxy | Tor, SOCKS |
| Command and Control | T1572 — Protocol Tunneling | Chisel, SoftEther |

**T1588.007 deserves separate notice.** This technique describes an adversary *acquiring AI capability* as a preparatory resource, on the same footing as buying infrastructure or signing certificates. Seeing it in a real-world mapping, rather than only in the documentation, is a milestone worth recording.

## IOCs

> From CrowdStrike's report.

**C2 infrastructure (BLOODALCHEMY)**

```plaintext
nginxmysql[.]com           45.153.126[.]223
sqlserver-sqlite[.]com     89.46.234[.]50
www[.]ddduuuppp[.]com      192.121.47[.]20
```

**SHA-256 hashes**

```plaintext
49fe70419fd19b81ca8d730a6ef7c1ec71b681d60b390761070d42686c7d2068   GameHook binary renamed svchost.exe
3fb5a9fb3e2cb1d6208966bd4e55d10282d19c7250d827efa560463d60a12ee4   BLOODALCHEMY loader (graphics-hook-filter64.dll)
364251838edfcf5ef1107fa692da3c5bd9a540482449f645992830134bf807ab   BLOODALCHEMY payload (graphics-hook_config.bin)
```

**ARTEX markers (longer-lived than the C2)**

```plaintext
Command-line parameters:          strings "artx" or "artex"
Cloud roles:                      artx- prefix
Administration commands:          "art18-" comments
Directory:                        C:\ProgramData\artx5v\
```

**Other artifacts**

```plaintext
D:\ProjectX\Merman_New\          # internal path in the BLOODALCHEMY implant
svchost.exe                      # masquerade name of the GameHook binary
graphics-hook-filter64.dll       # sideloaded malicious DLL
graphics-hook_config.bin         # encrypted payload
Operator instructions in Chinese, timestamped in China Standard Time
PRC red-team slang for Falcon: 猎鹰
```

**Tooling to watch for**

```plaintext
Chisel, SoftEther, Tor, SOCKS proxies
JSP and PHP web shells
LAPS password reads
RBCD abuse for impersonation
```

## References

- CrowdStrike — *VAPOR PANDA Uses AI-Driven ARTEX to Scale a Global Espionage Campaign* (Oct 29, 2026): <https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/>
- BleepingComputer — *Chinese hackers use ARTEX AI in automated attacks against 200 orgs*: <https://www.bleepingcomputer.com/news/security/chinese-hackers-use-artex-ai-in-automated-attacks-against-200-orgs/>
- CyberScoop — *Microsoft pins ConnectWise supply-chain attack to China-state aligned threat group* (on Microsoft's Storm-0940 report, Oct 21, 2026): <https://cyberscoop.com/microsoft-attributes-connectwise-supply-chain-attack-china-storm-0940/>
- SlashID — *Storm-0940 and ARTEX: Agentic OAuth Abuse in the ConnectWise Compromise* (ARTEX architecture analysis): <https://www.slashid.dev/blog/storm-0940-artex/>
