Skip to main content
Back
IntrusionAWS
6 min read

ShadowRay 2.0 attacks on exposed Ray clusters

Victim
Unnamed organizations running internet-exposed Ray dashboards
When
Not stated; first dated evidence Nov 2, 2025Disclosed Nov 18, 2025
Entry
Ray dashboards open to the internet with no authentication; attackers submitted jobs to the unauthenticated Ray Jobs API (CVE-2023-48022, which Ray's maintainers dispute) to run code on every node.

Oligo identified the campaign in early November 2025 and published on November 18. Attackers found exposed Ray dashboards through callbacks, submitted jobs that ran on every live node, added cron, systemd and SSH-key persistence and opened reverse shells. They ran XMRig to mine Monero and Rigel on GPUs, took MySQL credentials from environment variables and configuration files, and used the clusters to infect other dashboards; whether any cloud credential or data was exfiltrated is not stated.

What happened

External Parties
Command and Mining Hosts
GitLab and GitHub
Callback Service
+ 1 more
Internet
Exposed Ray Clusters
+ cloud credentials
+ EKS roles
+ security tokens
Cloud Roles on the Nodes
Ray Dashboard and Jobs API
Head and Worker Nodes
Models, Code, Datasets

Replaying the ShadowRay 2.0 November 2025 intrusion

  1. 1

    Exposed dashboard

    From early Nov 2025; start not stated

    Ray dashboards open to the internet accept jobs on the Jobs API with no authentication; callbacks mark the exploitable ones.

  2. 2

    Takeover of every node

    Nov 2025; exact dates not stated

    A payload runs on every live node, plants persistence and opens reverse shells to command hosts.

  3. 3

    Credentials and data

    Nov 2025; exfiltration not stated

    Payloads search for AWS credentials and take MySQL credentials from the environment and configuration files; whether anything was exfiltrated is not stated.

  4. 4

    Mining and spread

    Nov 2025; exact dates not stated

    XMRig mines Monero on the nodes and Rigel runs on their GPUs, and compromised clusters scan for and infect other Ray dashboards.

Impact

CPUs and GPUs of exposed Ray clusters were hijacked for cryptomining, with XMRig mining Monero and Rigel running on GPUs, and attackers gained reverse shells, persistence and, in several instances, access to AI models, datasets, source code and database credentials; compromised clusters scanned for and infected other dashboards. Oligo reports security tokens and cloud credentials on compromised machines and searches for AWS credentials, and does not say any was exfiltrated. The number of victims is not stated.

Outcome

Where cloud-provider controls fit

The entry is an application that has no authentication and expects a closed network. A cloud network control that keeps the dashboard off the internet stops the chain at its first step. After that, only egress rules and role limits bound what a compromised cluster does. Oligo reports that the campaign was still active when it published.

Could this have been prevented?

In hindsight, not as a whole. 2 controls would have refused a step outright, and 6 controls would have limited one. 4 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.

2
Prevented
6
Contained
4
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
Command and Mining Hosts
GitLab and GitHub
Callback Service
+ 1 more
Internet
Exposed Ray Clusters
+ cloud credentials
+ EKS roles
+ security tokens
Cloud Roles on the Nodes
Ray Dashboard and Jobs API
Head and Worker Nodes
Models, Code, Datasets

Replaying the ShadowRay 2.0 November 2025 intrusion

1 · Exposed dashboard

2 Prevented1 Contained3 on other clouds

2 · Takeover of every node

1 Contained2 Not covered3 on other clouds

3 · Credentials and data

2 Contained2 Not covered3 on other clouds

4 · Mining and spread

2 Contained6 on other clouds

Key facts

  • Oligo identified the active campaign in early November 2025 and documented payload changes on November 2 and 4, 2025; it says the operation could have been active since September 2024
  • GitLab removed the attacker account on November 5, 2025, the attackers moved to GitHub on November 10, GitHub blocked the account on November 17 and a new account was running within about 2 hours
  • Oligo's page gives both over 200,000 and more than 230,000 internet-exposed Ray servers, from Shodan
  • One payload used Ray's NodeAffinitySchedulingStrategy to run commands on every live node, and one compromised cluster had more than 1,000 nodes
  • Reverse shells connected to hosts in eight countries, and one stayed open for about 22 hours
  • One NFS mount held a 240 GB archive of source code, models and datasets; whether any data was exfiltrated is not stated

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.
    ShadowRay 2.0: Attackers Turn AI Against Itself in Global Campaign that Hijacks AI Into Self-Propagating BotnetOligo Security · Nov 18, 2025 · Researchtimeline, initial access, attack stages, cloud credential statements, persistence, counts, attribution, hedges, mitigations
  2. 2.
    Self-replicating botnet attacks Ray clustersThe Register · Nov 18, 2025 · Newsvendor position, exposure count as reported, start date as reported, takedown dates, cloud credential statement as reported
  3. 3.
    SecurityRay project documentation · undated · Vendor reporttrust model, dashboard and jobs exposure, deployment guidance, token authentication

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