Skip to main content
Back
IntrusionAzure · AWS · Kubernetes
7 min read

TeamPCP (PCPcat) worm through exposed Docker and Kubernetes APIs on cloud-hosted servers

Victim
Unnamed operators of internet-exposed servers, most of them on Azure and AWS
When
December 2025; actor channel active from Nov 17, 2025Disclosed Feb 5, 2026
Entry
A worm reached internet-exposed, unauthenticated Docker Engine APIs, Ray dashboards and React or Next.js applications on cloud-hosted servers, then ran a shared installer for persistence, tunneling, scanning, credential theft and, inside a pod, Kubernetes propagation.

From December 2025, a worm tracked as TeamPCP (PCPcat) scanned cloud provider address ranges for unauthenticated Docker Engine APIs and Ray dashboards and for vulnerable React and Next.js applications, then ran an installer that registered restart services, opened tunnels and fetched a miner. Flare fingerprinted 185 servers compromised through the Docker API, most of them on Azure and AWS, and read in the actor's script a Kubernetes path that spreads through a pod's mounted token and plants a privileged DaemonSet, with no cluster named. No source reports a stolen cloud credential used against a provider's API.

What happened

External Parties
Actor Staging and Control API
Research Honeypots
JobsGO (Claimed)
Internet
Internet-Exposed Cloud Servers
Kubernetes Clusters
Docker Engine API
Ray Dashboard
React or Next.js App
Installer and Services
Keys and Env Files
Pod Service Account Token
Cluster API Server
Other Pods
kube-system DaemonSet

Replaying the TeamPCP December 2025 intrusion

  1. 1

    Reaching exposed control planes

    December 2025; honeypot capture on December 14, 2025

    From cloud provider address lists, the worm reaches unauthenticated Docker Engine APIs and starts a container that runs its installer; its code also submits jobs to exposed Ray dashboards.

  2. 2

    Foothold and persistence on the host

    December 2025

    The installer registers restart services, opens frps and gost tunnels and fetches scanners and a miner from the actor's staging server.

  3. 3

    Web application credential path

    December 2025, from the actor's exploit script

    A crafted request to a vulnerable React or Next.js application runs commands, and the script posts the server's keys and environment files to the control API.

  4. 4

    Kubernetes propagation

    December 2025, from the actor's Kubernetes script

    Inside a pod, the installer uses the mounted token to run itself in every pod it can reach and creates a privileged DaemonSet in kube-system on every node.

  5. 5

    Mining and data release

    December 2025 to January 8, 2026

    A miner runs on the infected hosts, and the actor's channels release personal data dumps and claim a recruitment platform breach.

Impact

Flare fingerprinted 185 servers compromised through the Docker API in one phase, per its analysis of February 5, 2026; each host became a scanning and relaying node that mined cryptocurrency, and the web exploit script collected credential files. The actor's channels released personal data dumps of about 27 GB in 19 archives. The Ray, web application and Kubernetes paths are described from the actor's code with no named victim, and no source reports a stolen cloud credential used against a provider's API.

Outcome

Exposed management APIs carried the worm

The worm needed no credential: it reached management interfaces that answered anyone on the internet. You close those ports with inbound network rules, and you deny outbound internet traffic where a workload needs none, which refuses both the entry and the tunnels. Once code runs on a host, its files, services and compute are outside what a cloud control guards, and the web application flaw is the owner's to patch. No source reports a stolen cloud credential used against a provider's API.

Could this have been prevented?

In hindsight, not as a whole. 6 controls would have refused a step outright, and 7 controls would have limited one. 6 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

6
Prevented
7
Contained
6
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

External Parties
Actor Staging and Control API
Research Honeypots
JobsGO (Claimed)
Internet
Internet-Exposed Cloud Servers
Kubernetes Clusters
Docker Engine API
Ray Dashboard
React or Next.js App
Installer and Services
Keys and Env Files
Pod Service Account Token
Cluster API Server
Other Pods
kube-system DaemonSet

Replaying the TeamPCP December 2025 intrusion

1 · Reaching exposed control planes

4 Prevented4 on other clouds

2 · Foothold and persistence on the host

2 Prevented1 Not covered2 on other clouds

3 · Web application credential path

2 Contained2 Not covered2 on other clouds

4 · Kubernetes propagation

5 Contained1 Not covered2 on other clouds

5 · Mining and data release

2 Not covered

Key facts

  • Flare fingerprinted 185 servers compromised through exposed Docker APIs in one phase of the campaign, 182 of them tied to one control address, in its analysis of February 5, 2026
  • The actor's open statistics endpoint showed 91,505 targets scanned and 59,128 results collected on December 14, 2025, per Beelzebub
  • The group's own dashboard claimed more than 59,000 servers compromised in under 48 hours in December 2025; Flare presented it as the group's claim
  • Flare put 61% of compromised servers on Azure and 36% on AWS, without a sample size or a tie to the 185
  • A runtime capture showed the XMRig miner with 15 workers, and Flare said the mining yielded little
  • ShellForce claimed about 2.3 million JobsGO candidate records on January 8, 2026, an unverified claim with no statement by JobsGO found

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    Threat Alert: TeamPCP, An Emerging Force in the Cloud Native and Ransomware LandscapeFlare · Feb 5, 2026 · Researchaliases, Telegram dates, campaign peak, Docker phase on honeypots, compromised server count, provider split, installer behavior, Kubernetes script behavior, React exploit script, credential collection, scanner behavior, mining, tunnels, Sliver, leaked data, JobsGO claim, defensive advice
  2. 2.
    Operation PCPcat: Hunting a Next.js Credential Stealer That's Already Compromised 59K ServersBeelzebub · undated · ResearchDocker API attempt on a honeypot, installer, tunnels, system services, credential files targeted, control API, statistics endpoint figures, CVE naming
  3. 3.
    Unmasking TeamPCP, King of Software Supply Chain AttacksFlare · Aug 27, 2026 · ResearchPCPcat as first campaign, 185 Docker compromises, hedge on the 59,000 figure, link to later supply chain attacks
  4. 4.
    Critical Security Vulnerability in React Server ComponentsReact · Dec 3, 2025 · Advisoryvulnerability description, CVE id, affected and fixed versions, Next.js affected

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns