Summary
Knowns through version 0.33.0 contains a server-side request forgery vulnerability in the POST /api/embedding-models/test endpoint. The endpoint sends outbound requests to caller-supplied destinations without sufficient validation, allowing attackers to probe internal hosts and cloud metadata services through transport error messages.
Affected context
The vulnerability occurs in deployments that expose the embedding-model testing endpoint and allow request data to influence outbound connection destinations. It becomes risky when the server can reach internal networks, local services, or cloud metadata endpoints and returns transport or connectivity details to the caller.
Exploit preconditions
- Attack surface: The POST /api/embedding-models/test HTTP endpoint
- Required access: Ability to submit requests to the endpoint, subject to the application's access controls
- Trigger point: Processing and dispatch of a caller-supplied embedding-model destination
- Why exploitation is possible: The server initiates outbound connections without adequately restricting destination hosts, while transport errors can disclose whether network locations are reachable
Why this happens
The underlying bug is an insufficiently constrained server-side network request. User-controlled destination data is passed into an outbound request mechanism without strong allowlisting, destination validation, or protection against access to private and metadata address spaces. This behavior maps to CWE-918: Server-Side Request Forgery, in which an application is induced to make network requests on an attacker's behalf.
Potential impact
- Attackers may enumerate internal hosts and services
- Transport errors could disclose network reachability information
- The application may access cloud instance metadata endpoints
- Internal service details or metadata could be exposed
- Network reconnaissance could support follow-on attacks
General mitigation guidance
Mitigation generally focuses on reducing trust in caller-supplied destinations and limiting the server's network reach.
Common defensive patterns include:
- Restricting outbound requests to an explicit allowlist of approved schemes, hosts, and ports
- Blocking loopback, link-local, private, reserved, and cloud metadata address ranges
- Resolving and validating destinations before connection, including protections against DNS rebinding
- Using secure-by-default egress policies and isolated network execution for model testing
- Returning generic error responses that do not reveal internal network reachability
- Running the service with minimal network and cloud identity privileges
How teams detect issues like this
Organizations commonly detect SSRF risks through:
- Static analysis for outbound request functions that consume request-derived URLs or hostnames
- Code review checks for missing URL scheme, hostname, port, redirect, and IP-range validation
- Dynamic testing with controlled internal-address and metadata-service test cases
- Egress monitoring for unexpected connections to private, loopback, link-local, or metadata destinations
- API policy enforcement that restricts destination fields and normalizes error responses
Automated security review tools may help identify request flows from untrusted input to network clients.

