CVE-2023-48022: Unauthenticated Remote Code Execution in Ray Jobs API (ShadowRay)
- Vendor
- Anyscale
CVE-2023-48022 allows remote attackers to execute arbitrary code through Ray's unauthenticated job submission API, including in Ray 2.6.3 and 2.8.0. A client with network access to an unprotected submission interface can submit a job whose entrypoint specifies an attacker-controlled shell command. Execution occurs in the Ray job or worker security context, and Ray scheduling capabilities can distribute malicious workloads across cluster nodes. The vendor disputes the vulnerability classification because Ray is intended to execute trusted code within a strictly controlled network environment. Ray 2.52.0 and later offer optional token authentication; upgrading alone does not establish that this protection is enabled.
Impact, mitigation & remediation
What it means and what to do now: patch path, interim mitigations, and the assume-compromise checklist.
Impact
What an attacker gets, and what they’ve been doing with it.
Successful exploitation compromises the confidentiality, integrity, and availability of resources accessible to the Ray execution account. Attackers can access mounted datasets, model weights, source code, cloud credentials, service-account tokens, and other workload secrets, and consume CPU and GPU resources. Cluster scheduling can extend malicious execution to additional nodes, including nodes without direct internet exposure. Observed exploitation includes cryptomining, reverse shells, credential and data theft, cross-cluster propagation, and DDoS activity. System-level control depends on the execution account's privileges or subsequent privilege escalation.
Mitigation
If you can’t patch tonight, do this now.
Restrict Ray dashboard access, normally on TCP port 8265, and job-agent interfaces using firewall rules, private network segments, cloud security groups, and authenticated access controls. Permit connections only from trusted clients and cluster peers; protecting only the dashboard is insufficient if a job agent remains exposed. Run workloads with least privilege and limit accessible credentials and mounted sensitive data. Monitor job submissions, unexpected shell processes, outbound connections, persistence changes, and unexplained CPU or GPU use. Validate exposure using Ray Open Ports Checker or equivalent network checks.
Remediation
Patch, then assume compromise.
No conventional patch or fixed release is identified for this disputed issue. Deploy Ray only within its intended strictly controlled network environment and remove untrusted access to all job submission interfaces. Where appropriate, upgrade to Ray 2.52.0 or later and explicitly configure optional token authentication. Treat authentication as an additional control rather than a replacement for network isolation, and verify that both dashboard and job-agent interfaces are protected.
Exploits
3 exploits validated by Mallory's analysis.
Affected products & vendors
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
References
Frequently asked questions
- Is CVE-2023-48022 exploited in the wild?
- Yes. As of Nov 19, 2026, Mallory has reports of exploitation in the wild.
- Is there a public exploit for CVE-2023-48022?
- Yes. Mallory tracks 3 public exploits or proofs of concept for CVE-2023-48022, the earliest disclosed Nov 15, 2023.
- Is there a patch for CVE-2023-48022?
- No conventional patch or fixed release is identified for this disputed issue. Deploy Ray only within its intended strictly controlled network environment and remove untrusted access to all job submission interfaces.
- What products are affected by CVE-2023-48022?
- Anyscale Ray.
- How severe is CVE-2023-48022?
- CVE-2023-48022 has a CVSS v3.1 score of 9.8 (Critical) and an EPSS score of 83.9% (100th percentile) as of Nov 19, 2026.
Recent activity
32 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.