CVE-2023-46737: Possible endless data attack from attacker-controlled registry

Summary

Cosign is susceptible to a denial of service by an attacker controlled registry. An attacker who controls a remote registry can return a high number of attestations and/or signatures to Cosign and cause Cosign to enter a long loop resulting in an endless data attack. The root cause is that Cosign loops through all attestations fetched from the remote registry in pkg/cosign.FetchAttestations.

The attacker needs to compromise the registry or make a request to a registry they control. When doing so, the attacker must return a high number of attestations in the response to Cosign. The result will be that the attacker can cause Cosign to go into a long or infinite loop that will prevent other users from verifying their data. In Kyvernos case, an attacker whose privileges are limited to making requests to the cluster can make a request with an image reference to their own registry, trigger the infinite loop and deny other users from completing their admission requests. Alternatively, the attacker can obtain control of the registry used by an organization and return a high number of attestations instead the expected number of attestations.

The vulnerable loop in Cosign starts on line 154 below:

func FetchAttestations(se oci.SignedEntity, predicateType string) ([]AttestationPayload, error) {

atts, err := se.Attestations()

if err != nil {

return nil, fmt.Errorf("remote image: %w", err)

}

l, err := atts.Get()

if err != nil {

return nil, fmt.Errorf("fetching attestations: %w", err)

}

if len(l) == 0 {

return nil, errors.New(“found no attestations”)

}

attestations := make([]AttestationPayload, 0, len(l))

var attMu sync.Mutex

var g errgroup.Group

g.SetLimit(runtime.NumCPU())

for _, att := range l {

att := att

g.Go(func() error {

rawPayload, err := att.Payload()

if err != nil {

return fmt.Errorf("fetching payload: %w", err)

}

var payload AttestationPayload

if err := json.Unmarshal(rawPayload, &payload); err != nil {

return fmt.Errorf("unmarshaling payload: %w", err)

}

if predicateType != “” {

var decodedPayload []byte

decodedPayload, err = base64.StdEncoding.DecodeString(payload.PayLoad)

if err != nil {

return fmt.Errorf("decoding payload: %w", err)

}

var statement in_toto.Statement

if err := json.Unmarshal(decodedPayload, &statement); err != nil {

return fmt.Errorf("unmarshaling statement: %w", err)

}

if statement.PredicateType != predicateType {

return nil

}

}

attMu.Lock()

defer attMu.Unlock()

attestations = append(attestations, payload)

return nil

})

}

if err := g.Wait(); err != nil {

return nil, err

}

if len(attestations) == 0 && predicateType != “” {

return nil, fmt.Errorf("no attestations with predicate type ‘%s’ found", predicateType)

}

return attestations, nil

}

The l slice is controllable by an attacker who controls the remote registry.

Many cloud-native projects consider the remote registry to be untrusted, including Crossplane, Notary and Kyverno. We consider the same to be the case for Cosign, since users are not in control of whether the registry returns the expected data.

TUF’s security model labels this type of vulnerability an "Endless data attack", but an attacker could use this as a type of rollback attack, in case the user attempts to deploy a patched version of a vulnerable image; The attacker could prevent this upgrade by causing Cosign to get stuck in an infinite loop and never complete.

Mitigation

The issue can be mitigated rather simply by setting a limit to the limit of attestations that Cosign will loop through. The limit does not need to be high to be within the vast majority of use cases and still prevent the endless data attack.