CVE-2025-62593 is a critical-severity vulnerability. AttackerView checks Ray sites for it with a live exploit check.
Detects Ray clusters whose dashboard API accepts shell commands from a web page. Ray ships the dashboard HTTP API with no authentication by design, so POST /api/jobs/ runs any entrypoint command for anyone who can reach port 8265. Before 2.52.0 the only thing keeping a browser out was a check that the User-Agent header does not start with Mozilla. The fetch specification lets a page set that header, and Firefox and Safari implement it, so a page simply sends User-Agent: Other and walks through. Pair that with DNS rebinding, which moves the page onto the victim's own loopback or onto a Ray node inside the corporate network, and a malicious site or a malicious ad gets code execution on a developer machine or on an internal cluster the attacker cannot reach directly. The verifier first confirms the guard is live by watching a real Firefox User-Agent get refused with 405, then resends the identical request the way a browser does, keeping the Sec-Fetch headers a page cannot strip, and finally reads the job's own stdout back to show the target's shell computed an arithmetic result that was never sent to it. CVSS 9.4. Upgrade Ray to 2.52.0 or later, turn on the token authentication that release added, and keep the dashboard port off untrusted networks.
Is your app exploitable through CVE-2025-62593?
Scan your domain free