Unauthenticated SSRF reaches loopback and Docker host services in Portkey AI Gateway

Note: This data is for reference and cybersecurity research purposes only.CyStack advises users not to use this information for unlawful purposes.

What is CVE-2026-110719?

CVE-2026-110719 is a vulnerability classified as Server-Side Request Forgery, affecting gateway (affected versions: 1.14.0 – ≤ 1.15.2). This vulnerability is rated Medium, with a CVSS score of 6.9. Public exploit code or evidence is available for this vulnerability, but that does not confirm exploitation in the wild.

Overview

Original source data

Portkey AI Gateway 1.14.0 through 1.15.2 contains a server-side request forgery vulnerability that allows unauthenticated attackers to reach loopback and Docker host services because isValidCustomHost trusts localhost-style hosts when TRUSTED_CUSTOM_HOSTS is unset. Attackers can send /v1 requests with a crafted x-portkey-custom-host header, such as via the ollama provider, to make the gateway call internal services and return their responses.

Affected products and scope

  • Portkey-AI Portkey AI Gateway: affected from 1.14.0 through 1.15.2, using semver.
  • The normalized record sets default_status to unaffected outside the declared affected range, but it does not provide separate branch-by-branch validation for parallel release lines.
  • No exact fixed release appears in the supplied solution text or the advisory reviewed. The status of releases after the stated range should be verified against vendor release metadata.

Technical details

Portkey AI Gateway accepts a user-controlled x-portkey-custom-host header and uses it as the base URL for provider requests. When TRUSTED_CUSTOM_HOSTS is unset, isValidCustomHost trusts localhost, 127.0.0.1, ::1, host.docker.internal, and *.localhost; this trusted-host path returns success before private-IP and metadata checks are applied. The /v1/* routes invoke requestValidator, after which the provider context builds a URL from the custom host and passes the request through the gateway's fetch path. The Ollama provider does not require an API key for this path, so a remote unauthenticated attacker can cause the gateway to contact loopback or Docker-host services and relay their responses. The demonstrated scope centers on the default-trusted destinations; the public analysis does not claim RCE or unrestricted access to every internal network.

Exploitability

  • The issue is remotely reachable when a self-hosted gateway is accessible to the attacker.
  • No prior authentication or user interaction is required, and the described attack conditions are low complexity.
  • An attacker sends a /v1 request with a crafted x-portkey-custom-host header; the Ollama provider can be used because it does not require an API key.
  • Public exploit material includes source analysis and testing against a gateway owned by the researcher. That material does not demonstrate exploitation of a third-party deployment.
  • The supplied evidence does not establish exploitation in the wild.

Technical impact

  • Confidentiality: the gateway may read and return responses from HTTP services on loopback and the Docker host, including services bound only for local access.
  • Integrity: forwarded requests may affect internal data or configuration when the destination permits changes; the record does not identify a specific endpoint or confirm a successful modification.
  • Scope: the effect can extend beyond the gateway process to services reachable from its runtime environment.
  • Technical limit: the public analysis states that private IPv4, private IPv6, reserved ranges, metadata endpoints, and blocked TLDs are rejected for hosts outside the trusted set. It also notes that checking only the hostname string may be insufficient across DNS resolution or rebinding; that is a defensive concern identified by the analysis, not independent evidence of every possible exploit path.
  • The record contains no evidence of RCE or a specific availability impact.

Business impact

  • Internal data exposure: the gateway can contact HTTP services on loopback or the Docker host and return their responses to an external client. This may expose internal APIs or administrative interfaces that were not intended to be public.
  • Expanded reachability: a public gateway endpoint can become an intermediary for probing ports and services accessible only from the gateway's runtime environment.
  • Integrity risk: if an internal service accepts the forwarded request and permits state-changing operations, unauthorized changes may be possible. The record does not identify a specific target or confirm a successful change.
  • Limit: the available evidence focuses on reading or forwarding responses from loopback and Docker-host services; it does not confirm code execution or a specific service outage.

Remediation

  1. Apply the patch or fixed build indicated as available by the record, and verify the exact corrected release against Portkey-AI package or release metadata before deployment. The record does not specify an exact fixed release.
  2. Until that verification is complete, do not leave TRUSTED_CUSTOM_HOSTS unset. Configure a narrow, controlled allowlist containing only the custom hosts that are required, and do not trust loopback or Docker-host destinations by default.
  3. Enforce authentication and authorization before /v1/* routes, and consider disabling custom-host handling for public-facing deployments when the feature is not required.
  4. Strengthen SSRF defenses by validating the resolved IP immediately before connection, applying private, reserved, and metadata blocks to trusted hosts as well, and preventing automatic redirects to unvalidated destinations.
  5. After the change, retest every route that processes x-portkey-custom-host, including the Ollama provider, to ensure the allowlist is applied consistently.

Detection

  1. Inventory every Portkey AI Gateway deployment and compare its package, image, or running commit with the affected range in the record.
  2. Inspect runtime configuration, especially whether TRUSTED_CUSTOM_HOSTS is unset. An unset variable is a material condition for this issue.
  3. If request logs retain headers, review /v1 requests containing x-portkey-custom-host and requests selecting the ollama provider. This is a precautionary review based on the documented attack path, not a confirmed IOC.
  4. Review outbound connections from the gateway to localhost, 127.0.0.1, ::1, host.docker.internal, and hostnames ending in .localhost, along with internal responses returned to clients.
  5. Correlate unusual gateway-originated internal requests with the corresponding /v1 requests. Lack of matching log evidence does not prove that a deployment is safe.
Sources (26)
Learn more

Run an in-depth assessment with complete web risk management

CyStack VulnScan continuously discovers assets, validates vulnerabilities, and helps security teams prioritize remediation across the organization.