diff --git a/packages/@n8n/nodes-langchain/nodes/embeddings/EmbeddingsAzureOpenAi/EmbeddingsAzureOpenAi.node.ts b/packages/@n8n/nodes-langchain/nodes/embeddings/EmbeddingsAzureOpenAi/EmbeddingsAzureOpenAi.node.ts
index d36b0e722df..2267867469a 100644
--- a/packages/@n8n/nodes-langchain/nodes/embeddings/EmbeddingsAzureOpenAi/EmbeddingsAzureOpenAi.node.ts
+++ b/packages/@n8n/nodes-langchain/nodes/embeddings/EmbeddingsAzureOpenAi/EmbeddingsAzureOpenAi.node.ts
@@ -1,3 +1,4 @@
+import { getBearerTokenProvider } from '@azure/identity';
 import { AzureOpenAIEmbeddings, OpenAIEmbeddings } from '@langchain/openai';
 import {
 	getProxyAgent,
@@ -6,14 +7,63 @@ import {
 	getConnectionHintNoticeField,
 } from '@n8n/ai-utilities';
 import {
+	assertUrlAllowed,
+	getCredentialAllowedDomains,
 	NodeConnectionTypes,
 	NodeOperationError,
+	type ICredentialDataDecryptedObject,
+	type INode,
 	type INodeType,
 	type INodeTypeDescription,
 	type ISupplyDataFunctions,
 	type SupplyData,
 } from 'n8n-workflow';
 
+import { N8nOAuth2TokenCredential } from '../../llms/LmChatAzureOpenAi/credentials/N8nOAuth2TokenCredential';
+import {
+	AuthenticationType,
+	AZURE_OPENAI_INFERENCE_SCOPE,
+} from '../../llms/LmChatAzureOpenAi/types';
+import type { AzureEntraCognitiveServicesOAuth2ApiCredential } from '../../llms/LmChatAzureOpenAi/types';
+
+const API_KEY_AUTH = AuthenticationType.ApiKey;
+const ENTRA_AUTH = AuthenticationType.EntraOAuth2;
+
+type AzureApiKeyCredential = {
+	apiKey: string;
+	resourceName?: string;
+	apiVersion?: string;
+	endpoint?: string;
+	endpointType?: 'classic' | 'foundry';
+	foundryEndpoint?: string;
+};
+
+/** The endpoint and deployment details both credential types have to supply. */
+type AzureTarget = Pick<
+	AzureApiKeyCredential,
+	'resourceName' | 'apiVersion' | 'endpoint' | 'endpointType' | 'foundryEndpoint'
+>;
+
+/**
+ * Applies the credential's domain restriction to a host the credential itself supplies.
+ *
+ * `'none'` is a no-op here. It means "do not use this credential in the HTTP Request node",
+ * and the editor writes it into every credential made through its OAuth flow, so honouring it
+ * would stop the node the credential belongs to. An explicit `'domains'` list still applies.
+ */
+function assertCredentialAllowsHost(
+	node: INode,
+	credentialData: ICredentialDataDecryptedObject,
+	url: string,
+): void {
+	const allowedDomains = getCredentialAllowedDomains({
+		node,
+		credentialData,
+		credentialOwnedSurface: true,
+	});
+	assertUrlAllowed({ url, allowedDomains, node });
+}
+
 export class EmbeddingsAzureOpenAi implements INodeType {
 	description: INodeTypeDescription = {
 		displayName: 'Microsoft Foundry Embeddings',
@@ -21,8 +71,22 @@ export class EmbeddingsAzureOpenAi implements INodeType {
 		icon: 'file:azure.svg',
 		credentials: [
 			{
-				name: 'azureOpenAiApi',
+				name: API_KEY_AUTH,
 				required: true,
+				displayOptions: {
+					show: {
+						authentication: [API_KEY_AUTH],
+					},
+				},
+			},
+			{
+				name: ENTRA_AUTH,
+				required: true,
+				displayOptions: {
+					show: {
+						authentication: [ENTRA_AUTH],
+					},
+				},
 			},
 		],
 		group: ['transform'],
@@ -60,6 +124,22 @@ export class EmbeddingsAzureOpenAi implements INodeType {
 		outputs: [NodeConnectionTypes.AiEmbedding],
 		outputNames: ['Embeddings'],
 		properties: [
+			{
+				displayName: 'Authentication',
+				name: 'authentication',
+				type: 'options',
+				default: API_KEY_AUTH,
+				options: [
+					{
+						name: 'API Key',
+						value: API_KEY_AUTH,
+					},
+					{
+						name: 'Azure Entra ID (OAuth2)',
+						value: ENTRA_AUTH,
+					},
+				],
+			},
 			getConnectionHintNoticeField([NodeConnectionTypes.AiVectorStore]),
 			{
 				displayName: 'Model (Deployment) Name',
@@ -136,14 +216,52 @@ export class EmbeddingsAzureOpenAi implements INodeType {
 
 	async supplyData(this: ISupplyDataFunctions, itemIndex: number): Promise<SupplyData> {
 		this.logger.debug('Supply data for embeddings');
-		const credentials = await this.getCredentials<{
-			apiKey: string;
-			resourceName?: string;
-			apiVersion?: string;
-			endpoint?: string;
-			endpointType?: 'classic' | 'foundry';
-			foundryEndpoint?: string;
-		}>('azureOpenAiApi');
+		const authentication = this.getNodeParameter('authentication', itemIndex, API_KEY_AUTH) as
+			| typeof API_KEY_AUTH
+			| typeof ENTRA_AUTH;
+
+		let target: AzureTarget;
+		// Typed credential shapes have no index signature, so they need widening
+		// to reach the restriction check. Same form as core's oauth.ts.
+		let credentialData: ICredentialDataDecryptedObject;
+		// Exactly one of these is set. The API key is a string the client keeps; the token provider
+		// is a function the client calls before each request, so a long run never reuses a stale
+		// token.
+		let apiKey: string | undefined;
+		let tokenProvider: (() => Promise<string>) | undefined;
+
+		if (authentication === ENTRA_AUTH) {
+			const credential =
+				await this.getCredentials<AzureEntraCognitiveServicesOAuth2ApiCredential>(ENTRA_AUTH);
+			const entraCredential = new N8nOAuth2TokenCredential(
+				this.getNode(),
+				credential,
+				undefined,
+				this.helpers.getSecureEgressFilter(),
+			);
+			const deployment = await entraCredential.getDeploymentDetails();
+
+			target = {
+				resourceName: deployment.resourceName,
+				apiVersion: deployment.apiVersion,
+				endpoint: deployment.endpoint,
+				endpointType: credential.endpointType,
+				foundryEndpoint: credential.foundryEndpoint,
+			};
+			tokenProvider = getBearerTokenProvider(entraCredential, AZURE_OPENAI_INFERENCE_SCOPE);
+			credentialData = credential as unknown as ICredentialDataDecryptedObject;
+		} else {
+			const credential = await this.getCredentials<AzureApiKeyCredential>(API_KEY_AUTH);
+			target = credential;
+			apiKey = credential.apiKey;
+			credentialData = credential as unknown as ICredentialDataDecryptedObject;
+		}
+
+		const credentialLabel =
+			authentication === ENTRA_AUTH
+				? 'Microsoft Foundry (Entra ID)'
+				: 'Microsoft Foundry (API Key)';
+
 		const modelName = this.getNodeParameter('model', itemIndex) as string;
 
 		const options = this.getNodeParameter('options', itemIndex, {}) as {
@@ -157,16 +275,19 @@ export class EmbeddingsAzureOpenAi implements INodeType {
 			options.timeout = undefined;
 		}
 
-		if (credentials.endpointType === 'foundry') {
-			const foundryURL = credentials.foundryEndpoint?.trim();
+		if (target.endpointType === 'foundry') {
+			const foundryURL = target.foundryEndpoint?.trim();
 			if (!foundryURL) {
 				throw new NodeOperationError(
 					this.getNode(),
-					'Foundry endpoint is missing in the selected Microsoft Foundry (API Key) credential.',
+					`Foundry endpoint is missing in the selected ${credentialLabel} credential.`,
 				);
 			}
+			assertCredentialAllowsHost(this.getNode(), credentialData, foundryURL);
+
 			const embeddings = new OpenAIEmbeddings({
-				apiKey: credentials.apiKey,
+				// The openai client accepts a `() => Promise<string>` here and calls it per request
+				apiKey: tokenProvider ?? apiKey,
 				model: modelName,
 				configuration: {
 					baseURL: foundryURL,
@@ -183,25 +304,40 @@ export class EmbeddingsAzureOpenAi implements INodeType {
 			};
 		}
 
+		// A classic deployment addresses <resource>.openai.azure.com/...?api-version=. An empty
+		// value for either reaches Azure as a malformed URL, so say which field is missing.
+		if (!target.endpoint?.trim() && !target.resourceName?.trim()) {
+			throw new NodeOperationError(
+				this.getNode(),
+				`Resource Name is missing in the selected ${credentialLabel} credential.`,
+			);
+		}
+		if (!target.apiVersion?.trim()) {
+			throw new NodeOperationError(
+				this.getNode(),
+				`API Version is missing in the selected ${credentialLabel} credential.`,
+			);
+		}
+
+		// `||` rather than `??`, so an endpoint that is set but empty also falls back.
+		const dialledHost = target.endpoint || `https://${target.resourceName}.openai.azure.com`;
+		assertCredentialAllowsHost(this.getNode(), credentialData, dialledHost);
+
 		const embeddings = new AzureOpenAIEmbeddings({
 			azureOpenAIApiDeploymentName: modelName,
 			// instance name only needed to set base url
-			azureOpenAIApiInstanceName: !credentials.endpoint ? credentials.resourceName : undefined,
-			azureOpenAIApiKey: credentials.apiKey,
-			azureOpenAIApiVersion: credentials.apiVersion,
+			azureOpenAIApiInstanceName: !target.endpoint ? target.resourceName : undefined,
+			azureOpenAIApiKey: apiKey,
+			azureADTokenProvider: tokenProvider,
+			azureOpenAIApiVersion: target.apiVersion,
 			// azureOpenAIEndpoint and configuration.baseURL are both ignored here
 			// only setting azureOpenAIBasePath worked
-			azureOpenAIBasePath: credentials.endpoint
-				? `${credentials.endpoint}/openai/deployments`
-				: undefined,
+			azureOpenAIBasePath: target.endpoint ? `${target.endpoint}/openai/deployments` : undefined,
 			configuration: {
 				fetch: aiClientFetch,
 				fetchOptions: {
-					dispatcher: getProxyAgent(
-						credentials.endpoint ?? `https://${credentials.resourceName}.openai.azure.com`,
-						{},
-						this.helpers.getSecureEgressFilter(),
-					),
+					// Resolve the proxy against the host LangChain dials so NO_PROXY applies to it.
+					dispatcher: getProxyAgent(dialledHost, {}, this.helpers.getSecureEgressFilter()),
 				},
 			},
 			...options,
diff --git a/packages/@n8n/nodes-langchain/nodes/embeddings/test/EmbeddingsAzureOpenAi.test.ts b/packages/@n8n/nodes-langchain/nodes/embeddings/test/EmbeddingsAzureOpenAi.test.ts
index 4d6ecada721..7a60f970047 100644
--- a/packages/@n8n/nodes-langchain/nodes/embeddings/test/EmbeddingsAzureOpenAi.test.ts
+++ b/packages/@n8n/nodes-langchain/nodes/embeddings/test/EmbeddingsAzureOpenAi.test.ts
@@ -1,11 +1,44 @@
 import { AzureOpenAIEmbeddings, OpenAIEmbeddings } from '@langchain/openai';
-import { aiClientFetch } from '@n8n/ai-utilities';
+import { aiClientFetch, getProxyAgent } from '@n8n/ai-utilities';
 import { createMockExecuteFunction } from 'n8n-nodes-base/test/nodes/Helpers';
 import type { INode, ISupplyDataFunctions } from 'n8n-workflow';
 import type { Mocked } from 'vitest';
 
 import { EmbeddingsAzureOpenAi } from '../EmbeddingsAzureOpenAi/EmbeddingsAzureOpenAi.node';
 
+const tokenCredentialSpy = vi.fn();
+
+// Records the constructor arguments but still derives the deployment from the credential it was
+// given, so the other tests keep asserting real mapping rather than a fixed stub.
+vi.mock('../../llms/LmChatAzureOpenAi/credentials/N8nOAuth2TokenCredential', () => ({
+	N8nOAuth2TokenCredential: class N8nOAuth2TokenCredentialMock {
+		private credential: Record<string, unknown> = {};
+
+		constructor(...args: unknown[]) {
+			tokenCredentialSpy(...args);
+			this.credential = (args[1] ?? {}) as Record<string, unknown>;
+		}
+
+		getDeploymentDetails = async () => {
+			const c = this.credential;
+			if (c.endpointType === 'foundry') {
+				return {
+					apiVersion: c.apiVersion ?? '',
+					endpoint: c.foundryEndpoint,
+					resourceName: c.resourceName ?? '',
+					endpointType: 'foundry' as const,
+					foundryEndpoint: c.foundryEndpoint,
+				};
+			}
+			return {
+				apiVersion: c.apiVersion ?? '',
+				endpoint: c.endpoint,
+				resourceName: c.resourceName ?? '',
+			};
+		};
+	},
+}));
+
 vi.mock('@langchain/openai');
 
 class MockProxyAgent {}
@@ -126,6 +159,86 @@ describe('AzureOpenAIEmbeddings', () => {
 			);
 		});
 
+		describe('credential domain restrictions', () => {
+			const restricted = {
+				allowedHttpRequestDomains: 'domains',
+				allowedDomains: 'allowed.example.com',
+			};
+
+			it.each([
+				[
+					'a Foundry endpoint',
+					{ endpointType: 'foundry', foundryEndpoint: 'https://evil.example.com/openai/v1' },
+				],
+				['a classic endpoint', { endpoint: 'https://evil.example.com', apiVersion: 'v1' }],
+				['a resource-derived classic host', { resourceName: 'evil', apiVersion: 'v1' }],
+			])('refuses to send the token to %s outside the allowed domains', async (_name, target) => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue({
+					apiKey: 'test-api-key',
+					...restricted,
+					...target,
+				});
+				mockContext.getNodeParameter = vi.fn().mockImplementation((paramName: string) => {
+					if (paramName === 'model') return 'text-embedding-3-large';
+					if (paramName === 'options') return {};
+					return undefined;
+				});
+
+				await expect(embeddingsAzureOpenAi.supplyData.call(mockContext, 0)).rejects.toThrow(
+					'Domain not allowed',
+				);
+			});
+
+			// 'none' means "do not use this credential in the HTTP Request node". The editor
+			// writes it into every credential made through its OAuth flow, so blocking on it
+			// here would stop the node the credential belongs to.
+			it.each([
+				[
+					'a Foundry endpoint',
+					{
+						endpointType: 'foundry',
+						foundryEndpoint: 'https://test.services.ai.azure.com/openai/v1',
+					},
+				],
+				[
+					'a classic endpoint',
+					{ endpoint: 'https://test-resource-name.openai.azure.com', apiVersion: 'v1' },
+				],
+			])('still reaches %s when the restriction is None', async (_name, target) => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue({
+					apiKey: 'test-api-key',
+					allowedHttpRequestDomains: 'none',
+					...target,
+				});
+				mockContext.getNodeParameter = vi.fn().mockImplementation((paramName: string) => {
+					if (paramName === 'model') return 'text-embedding-3-large';
+					if (paramName === 'options') return {};
+					return undefined;
+				});
+
+				await expect(embeddingsAzureOpenAi.supplyData.call(mockContext, 0)).resolves.toBeDefined();
+			});
+
+			it('allows a host the credential permits', async () => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue({
+					apiKey: 'test-api-key',
+					...restricted,
+					endpointType: 'foundry',
+					foundryEndpoint: 'https://allowed.example.com/openai/v1',
+				});
+				mockContext.getNodeParameter = vi.fn().mockImplementation((paramName: string) => {
+					if (paramName === 'model') return 'text-embedding-3-large';
+					if (paramName === 'options') return {};
+					return undefined;
+				});
+
+				await expect(embeddingsAzureOpenAi.supplyData.call(mockContext, 0)).resolves.toBeDefined();
+			});
+		});
+
 		it('should use OpenAIEmbeddings against the Foundry base URL', async () => {
 			const mockContext = setupMockContext();
 			const MockedOpenAIEmbeddings = vi.mocked(OpenAIEmbeddings);
@@ -157,6 +270,248 @@ describe('AzureOpenAIEmbeddings', () => {
 			expect(MockedAzureOpenAIEmbeddings).not.toHaveBeenCalled();
 		});
 
+		describe('with Entra ID', () => {
+			const entraCredential = {
+				clientId: 'client-id',
+				clientSecret: 'client-secret',
+				accessTokenUrl: 'https://login.microsoftonline.com/tenant/oauth2/token',
+				authentication: 'body',
+				scope: '',
+				tenantId: 'tenant',
+			};
+
+			const selectEntra = (context: Mocked<ISupplyDataFunctions>) => {
+				context.getNodeParameter = vi.fn().mockImplementation((paramName: string) => {
+					if (paramName === 'authentication') return 'azureEntraCognitiveServicesOAuth2Api';
+					if (paramName === 'model') return 'text-embedding-3-large';
+					if (paramName === 'options') return {};
+					return undefined;
+				});
+			};
+
+			// The Entra branch is the one this PR adds the egress check to, so it needs its own
+			// cases. The API-key copies cannot cover it: they dial a different code path.
+			describe('credential domain restrictions', () => {
+				const restricted = {
+					allowedHttpRequestDomains: 'domains',
+					allowedDomains: 'allowed.example.com',
+				};
+
+				it.each([
+					[
+						'a Foundry endpoint',
+						{ endpointType: 'foundry', foundryEndpoint: 'https://evil.example.com/openai/v1' },
+					],
+					['a classic endpoint', { endpoint: 'https://evil.example.com', apiVersion: 'v1' }],
+					['a resource-derived classic host', { resourceName: 'evil', apiVersion: 'v1' }],
+				])('refuses to send the token to %s outside the allowed domains', async (_name, target) => {
+					const mockContext = setupMockContext();
+					mockContext.getCredentials.mockResolvedValue({
+						...entraCredential,
+						...restricted,
+						...target,
+					});
+					selectEntra(mockContext);
+
+					await expect(embeddingsAzureOpenAi.supplyData.call(mockContext, 0)).rejects.toThrow(
+						'Domain not allowed',
+					);
+				});
+
+				// 'none' means "do not use this credential in the HTTP Request node". The editor
+				// writes it into every credential made through its OAuth flow, so blocking on it
+				// here would stop the node the credential belongs to.
+				it.each([
+					[
+						'a Foundry endpoint',
+						{
+							endpointType: 'foundry',
+							foundryEndpoint: 'https://test.services.ai.azure.com/openai/v1',
+						},
+					],
+					[
+						'a classic endpoint',
+						{ endpoint: 'https://test-resource-name.openai.azure.com', apiVersion: 'v1' },
+					],
+				])('still reaches %s when the restriction is None', async (_name, target) => {
+					const mockContext = setupMockContext();
+					mockContext.getCredentials.mockResolvedValue({
+						...entraCredential,
+						allowedHttpRequestDomains: 'none',
+						...target,
+					});
+					selectEntra(mockContext);
+
+					await expect(
+						embeddingsAzureOpenAi.supplyData.call(mockContext, 0),
+					).resolves.toBeDefined();
+				});
+
+				it('allows a host the credential permits', async () => {
+					const mockContext = setupMockContext();
+					mockContext.getCredentials.mockResolvedValue({
+						...entraCredential,
+						...restricted,
+						endpointType: 'foundry',
+						foundryEndpoint: 'https://allowed.example.com/openai/v1',
+					});
+					selectEntra(mockContext);
+
+					await expect(
+						embeddingsAzureOpenAi.supplyData.call(mockContext, 0),
+					).resolves.toBeDefined();
+				});
+			});
+
+			it('should pass a token provider to a classic deployment, and no API key', async () => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue({
+					...entraCredential,
+					endpoint: 'https://test-resource-name.openai.azure.com',
+					apiVersion: 'v1',
+				});
+				selectEntra(mockContext);
+
+				await embeddingsAzureOpenAi.supplyData.call(mockContext, 0);
+
+				expect(MockedAzureOpenAIEmbeddings).toHaveBeenCalledWith(
+					expect.objectContaining({
+						azureADTokenProvider: expect.any(Function),
+						azureOpenAIApiKey: undefined,
+						azureOpenAIBasePath: 'https://test-resource-name.openai.azure.com/openai/deployments',
+					}),
+				);
+			});
+
+			// A function, not an awaited string: the client calls it per request, so a long run never
+			// sends a token that has since expired.
+			it('should pass a token provider as the Foundry apiKey, not a resolved string', async () => {
+				const mockContext = setupMockContext();
+				const MockedOpenAIEmbeddings = vi.mocked(OpenAIEmbeddings);
+				mockContext.getCredentials.mockResolvedValue({
+					...entraCredential,
+					endpointType: 'foundry',
+					foundryEndpoint: 'https://test.services.ai.azure.com/openai/v1',
+				});
+				selectEntra(mockContext);
+
+				await embeddingsAzureOpenAi.supplyData.call(mockContext, 0);
+
+				const config = MockedOpenAIEmbeddings.mock.calls[0][0];
+				expect(typeof config?.apiKey).toBe('function');
+				expect(config?.model).toBe('text-embedding-3-large');
+				expect(config?.configuration).toEqual(
+					expect.objectContaining({ baseURL: 'https://test.services.ai.azure.com/openai/v1' }),
+				);
+				expect(MockedAzureOpenAIEmbeddings).not.toHaveBeenCalled();
+			});
+
+			// The host has to come from the resource name whether the endpoint is absent or blank.
+			// The blank case is the one that needs `||`: an empty string is falsy but not nullish,
+			// so `??` would pass it straight through to the proxy.
+			it.each([
+				['absent', {}],
+				['blank', { endpoint: '' }],
+			])(
+				'should resolve the proxy against the resource host when the endpoint is %s',
+				async (_, endpoint) => {
+					const mockContext = setupMockContext();
+					mockContext.getCredentials.mockResolvedValue({
+						...entraCredential,
+						...endpoint,
+						resourceName: 'my-resource',
+						apiVersion: 'v1',
+					});
+					selectEntra(mockContext);
+
+					await embeddingsAzureOpenAi.supplyData.call(mockContext, 0);
+
+					expect(vi.mocked(getProxyAgent)).toHaveBeenCalledWith(
+						'https://my-resource.openai.azure.com',
+						expect.any(Object),
+						expect.any(Object),
+					);
+				},
+			);
+
+			// The mint posts the client secret to a stored URL, so it runs inside the egress policy,
+			// as the chat model node's does.
+			it('should hand the egress filter to the token credential', async () => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue({
+					...entraCredential,
+					resourceName: 'my-resource',
+					apiVersion: 'v1',
+				});
+				selectEntra(mockContext);
+
+				await embeddingsAzureOpenAi.supplyData.call(mockContext, 0);
+
+				expect(tokenCredentialSpy).toHaveBeenCalledWith(
+					expect.anything(),
+					expect.anything(),
+					undefined,
+					mockContext.helpers.getSecureEgressFilter(),
+				);
+			});
+
+			it('should read the Entra credential, not the API key one', async () => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue({
+					...entraCredential,
+					endpoint: 'https://test-resource-name.openai.azure.com',
+					apiVersion: 'v1',
+				});
+				selectEntra(mockContext);
+
+				await embeddingsAzureOpenAi.supplyData.call(mockContext, 0);
+
+				expect(mockContext.getCredentials).toHaveBeenCalledWith(
+					'azureEntraCognitiveServicesOAuth2Api',
+				);
+				expect(mockContext.getCredentials).not.toHaveBeenCalledWith('azureOpenAiApi');
+			});
+		});
+
+		// Existing nodes have no `authentication` parameter stored. The Workflow constructor fills
+		// this default into node.parameters, and getCredentials gates the API key credential on it,
+		// so changing it would point every saved node at a credential it does not have.
+		it('should default authentication to the API key, for nodes saved before the selector', () => {
+			const authentication = new EmbeddingsAzureOpenAi().description.properties.find(
+				(p) => p.name === 'authentication',
+			);
+
+			expect(authentication?.default).toBe('azureOpenAiApi');
+		});
+
+		it.each([
+			[
+				'no resource name or endpoint',
+				{ apiKey: 'test-api-key', apiVersion: 'v1' },
+				'Resource Name is missing in the selected Microsoft Foundry (API Key) credential.',
+			],
+			[
+				'no API version',
+				{ apiKey: 'test-api-key', resourceName: 'my-resource' },
+				'API Version is missing in the selected Microsoft Foundry (API Key) credential.',
+			],
+		])(
+			'should say which field is missing when a classic credential has %s',
+			async (_, credential, message) => {
+				const mockContext = setupMockContext();
+				mockContext.getCredentials.mockResolvedValue(credential);
+				mockContext.getNodeParameter = vi.fn().mockImplementation((paramName: string) => {
+					if (paramName === 'model') return 'text-embedding-3-large';
+					if (paramName === 'options') return {};
+					return undefined;
+				});
+
+				await expect(embeddingsAzureOpenAi.supplyData.call(mockContext, 0)).rejects.toThrow(
+					message,
+				);
+			},
+		);
+
 		it('should reject a Foundry credential that has no endpoint', async () => {
 			const mockContext = setupMockContext();
 
diff --git a/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/credentials/oauth2.ts b/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/credentials/oauth2.ts
index a8ea1db0c89..16c1a0931bc 100644
--- a/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/credentials/oauth2.ts
+++ b/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/credentials/oauth2.ts
@@ -8,9 +8,8 @@ import type {
 	AzureEntraCognitiveServicesOAuth2ApiCredential,
 	AzureOpenAIOAuth2ModelConfig,
 } from '../types';
-import { AZURE_OPENAI_INFERENCE_AUDIENCE } from '../types';
+import { AZURE_OPENAI_INFERENCE_SCOPE } from '../types';
 
-const AZURE_OPENAI_SCOPE = `${AZURE_OPENAI_INFERENCE_AUDIENCE}/.default`;
 /**
  * Creates Entra ID (OAuth2) authentication for Azure OpenAI
  */
@@ -32,7 +31,10 @@ export async function setupOAuth2Authentication(
 
 		// getBearerTokenProvider caches the token across calls. It requires a scope, but the
 		// audience comes from the credential above; the v1.0 endpoint reads `resource`, not `scope`.
-		const azureADTokenProvider = getBearerTokenProvider(entraTokenCredential, AZURE_OPENAI_SCOPE);
+		const azureADTokenProvider = getBearerTokenProvider(
+			entraTokenCredential,
+			AZURE_OPENAI_INFERENCE_SCOPE,
+		);
 
 		this.logger.debug('Successfully created Azure AD Token Provider.');
 
diff --git a/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/types.ts b/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/types.ts
index a7ef1e0f5d7..e572e1cd11f 100644
--- a/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/types.ts
+++ b/packages/@n8n/nodes-langchain/nodes/llms/LmChatAzureOpenAi/types.ts
@@ -68,6 +68,13 @@ export interface AzureOpenAIOAuth2ModelConfig extends AzureOpenAIBaseModelConfig
 /** Audience for the node's inference requests. */
 export const AZURE_OPENAI_INFERENCE_AUDIENCE = 'https://cognitiveservices.azure.com';
 
+/**
+ * The audience in the form `getBearerTokenProvider` takes. It is only required by that API:
+ * `N8nOAuth2TokenCredential` ignores the scopes it is handed and asks for the audience as the
+ * `resource` body parameter, which is what the v1.0 token endpoint reads.
+ */
+export const AZURE_OPENAI_INFERENCE_SCOPE = `${AZURE_OPENAI_INFERENCE_AUDIENCE}/.default`;
+
 /**
  * Audience for the Foundry deployments-list call. Each tenant's Entra ID app
  * registration grants access per audience, so this is requested only for the