CVE-2026-27482 affects the Ray AI compute engine, specifically versions 2.53.0 and earlier. The issue stems from the dashboard HTTP server's handling of HTTP methods: while POST and PUT requests from browsers are blocked based on origin, DELETE requests are not similarly restricted. Moreover, key DELETE endpoints lack authentication by default. If the dashboard or agent is configured to be reachable externally (for example, by setting --dashboard-host=0.0.0.0), an attacker with network access or leveraging DNS rebinding can issue DELETE requests that forcibly shut down the Serve component or delete active jobs. This results in a denial-of-service condition impacting availability. The vulnerability is categorized under CWE-396, indicating improper exception handling practices, which in this context relates to the generic catch declaration that may contribute to insufficient request validation. The CVSS v3.1 base score is 5.9 (medium severity), reflecting network attack vector, high attack complexity, no privileges required, required user interaction, unchanged scope, no confidentiality impact, low integrity impact, and high availability impact. No known exploits are currently reported in the wild. The recommended remediation is upgrading to Ray 2.54.0 or later, where these endpoints are properly secured. Until then, administrators should avoid exposing the dashboard publicly and implement network-level access controls.

The primary impact of this vulnerability is on availability, as attackers can remotely shut down critical Ray Serve components or delete running jobs without authentication or user interaction. This can disrupt AI compute workloads, causing service outages and loss of in-progress computations. Organizations relying on Ray for AI model serving or distributed computing may experience downtime, operational delays, and potential loss of productivity. Since the vulnerability does not affect confidentiality or significantly impact integrity, data breaches or unauthorized data modifications are unlikely. However, the ease of exploitation via network access or DNS rebinding increases risk, especially in environments where the dashboard is exposed to untrusted networks. This can affect cloud deployments, research labs, and enterprises using Ray in multi-tenant or less isolated network configurations.

1. Upgrade all Ray deployments to version 2.54.0 or later immediately to apply the official fix. 2. Restrict access to the Ray dashboard and agent interfaces by binding them to localhost or internal IP addresses rather than 0.0.0.0. 3. Implement network-level controls such as firewalls, VPNs, or zero-trust segmentation to limit dashboard access to trusted users and systems only. 4. Disable or restrict DELETE HTTP methods on the dashboard server using reverse proxies or web application firewalls if upgrading is not immediately possible. 5. Monitor network traffic for unusual DELETE requests targeting the dashboard endpoints. 6. Educate administrators on the risks of exposing management interfaces publicly and enforce secure configuration baselines. 7. Consider deploying DNS rebinding protections on client networks to prevent exploitation via malicious web pages.