update huggingface to fix unbounded resources bug (#5204)

When a model's owner (user/org) isn't in the token's fine-grained scopes
AND the model itself isn't directly scoped, the Permission can end up
blank, which is a proto violation.  So, move those cases to unbounded
resources.
This commit is contained in:
Dave Hubbard
2026-08-13 13:42:15 -04:00
committed by GitHub
parent 37ed218cc3
commit ee476c7e56
2 changed files with 148 additions and 13 deletions
@@ -91,9 +91,9 @@ func bakeUnfineGrainedBindings(allModels []Model, tokenJSON HFTokenJSON) []analy
return bindings
}
// finegrained scopes are grouped by org, user or model.
func bakefineGrainedBindings(allModels []Model, tokenJSON HFTokenJSON) []analyzers.Binding {
// this section will extract the relevant permissions for each entity and store them in a map
// bakefineGrainedBindings resolves each model's permission from the token's
// fine-grained scopes (grouped by org, user, or model).
func bakefineGrainedBindings(allModels []Model, tokenJSON HFTokenJSON) ([]analyzers.Binding, []analyzers.Resource) {
var nameToPermissions = make(map[string]analyzers.Permission)
for _, permission := range tokenJSON.Auth.AccessToken.FineGrained.Scoped {
privs := analyzers.Permission{
@@ -115,10 +115,9 @@ func bakefineGrainedBindings(allModels []Model, tokenJSON HFTokenJSON) []analyze
}
}
bindings := make([]analyzers.Binding, len(allModels))
for idx, model := range allModels {
// Add Read Privs to All Models
var bindings []analyzers.Binding
var unbounded []analyzers.Resource
for _, model := range allModels {
modelResource := analyzers.Resource{
Name: model.Name,
FullyQualifiedName: "huggingface.com/model/" + model.ID,
@@ -129,22 +128,27 @@ func bakefineGrainedBindings(allModels []Model, tokenJSON HFTokenJSON) []analyze
}
var perm analyzers.Permission
// get username/orgname for each model and apply those permissions
modelUsername := strings.Split(model.Name, "/")[0]
if permissions, ok := nameToPermissions[modelUsername]; ok {
perm = permissions
}
// override model permissions with repo-specific permissions
if permissions, ok := nameToPermissions[model.Name]; ok {
perm = permissions
}
bindings[idx] = analyzers.Binding{
// Models without an explicit fine-grained scope are visible to the
// token but have no bound permission; they belong in UnboundedResources.
if perm.Value == "" {
unbounded = append(unbounded, modelResource)
continue
}
bindings = append(bindings, analyzers.Binding{
Resource: modelResource,
Permission: perm,
}
})
}
return bindings
return bindings, unbounded
}
func bakeOrganizationBindings(tokenJSON HFTokenJSON) []analyzers.Binding {
@@ -259,7 +263,9 @@ func secretInfoToAnalyzerResult(info *SecretInfo) *analyzers.AnalyzerResult {
result.Bindings = make([]analyzers.Binding, 0)
if info.Token.Auth.AccessToken.Type == FINEGRAINED {
result.Bindings = append(result.Bindings, bakefineGrainedBindings(info.Models, info.Token)...)
fgBindings, fgUnbounded := bakefineGrainedBindings(info.Models, info.Token)
result.Bindings = append(result.Bindings, fgBindings...)
result.UnboundedResources = append(result.UnboundedResources, fgUnbounded...)
result.Bindings = append(result.Bindings, bakeOrganizationBindings(info.Token)...)
result.Bindings = append(result.Bindings, bakeUserBindings(info.Token)...)
} else {
@@ -118,3 +118,132 @@ func TestAnalyzer_Analyze(t *testing.T) {
})
}
}
// Test_secretInfoToAnalyzerResult_FineGrainedUnboundedResources verifies that
// models outside the token's fine-grained permission scopes land in
// UnboundedResources rather than producing bindings with empty Permission.Value.
func Test_secretInfoToAnalyzerResult_FineGrainedUnboundedResources(t *testing.T) {
info := &SecretInfo{
Token: HFTokenJSON{
Username: "testuser",
Name: "Test User",
Auth: struct {
AccessToken struct {
Name string `json:"displayName"`
Type string `json:"role"`
CreatedAt string `json:"createdAt"`
FineGrained struct {
Global []string `json:"global"`
Scoped []struct {
Entity struct {
Type string `json:"type"`
Name string `json:"name"`
ID string `json:"_id"`
} `json:"entity"`
Permissions []string `json:"permissions"`
} `json:"scoped"`
} `json:"fineGrained"`
}
}{
AccessToken: struct {
Name string `json:"displayName"`
Type string `json:"role"`
CreatedAt string `json:"createdAt"`
FineGrained struct {
Global []string `json:"global"`
Scoped []struct {
Entity struct {
Type string `json:"type"`
Name string `json:"name"`
ID string `json:"_id"`
} `json:"entity"`
Permissions []string `json:"permissions"`
} `json:"scoped"`
} `json:"fineGrained"`
}{
Name: "test-token",
Type: FINEGRAINED,
FineGrained: struct {
Global []string `json:"global"`
Scoped []struct {
Entity struct {
Type string `json:"type"`
Name string `json:"name"`
ID string `json:"_id"`
} `json:"entity"`
Permissions []string `json:"permissions"`
} `json:"scoped"`
}{
Scoped: []struct {
Entity struct {
Type string `json:"type"`
Name string `json:"name"`
ID string `json:"_id"`
} `json:"entity"`
Permissions []string `json:"permissions"`
}{
{
Entity: struct {
Type string `json:"type"`
Name string `json:"name"`
ID string `json:"_id"`
}{Type: "user", Name: "testuser"},
Permissions: []string{"repo.content.read"},
},
},
},
},
},
},
Models: []Model{
{Name: "testuser/scoped-model", ID: "abc123", Private: true},
{Name: "other-org/unscoped-model", ID: "def456", Private: false},
{Name: "another-org/also-unscoped", ID: "ghi789", Private: false},
},
}
result := secretInfoToAnalyzerResult(info)
// The scoped model should appear as a binding with a non-empty permission.
if len(result.Bindings) == 0 {
t.Fatal("expected at least one binding for the scoped model")
}
foundScoped := false
for _, b := range result.Bindings {
if b.Resource.Name == "testuser/scoped-model" {
foundScoped = true
if b.Permission.Value == "" {
t.Error("scoped model binding has empty Permission.Value")
}
}
if b.Permission.Value == "" {
t.Errorf("binding for %q has empty Permission.Value (would fail proto validation)", b.Resource.Name)
}
}
if !foundScoped {
t.Error("expected testuser/scoped-model in Bindings")
}
// Unscoped models should appear in UnboundedResources, not Bindings.
unscopedNames := map[string]bool{
"other-org/unscoped-model": false,
"another-org/also-unscoped": false,
}
for _, r := range result.UnboundedResources {
if _, want := unscopedNames[r.Name]; want {
unscopedNames[r.Name] = true
}
}
for name, found := range unscopedNames {
if !found {
t.Errorf("expected %q in UnboundedResources, but it was missing", name)
}
}
// Verify unscoped models are NOT in Bindings.
for _, b := range result.Bindings {
if _, isUnscoped := unscopedNames[b.Resource.Name]; isUnscoped {
t.Errorf("unscoped model %q should not appear in Bindings", b.Resource.Name)
}
}
}