How A2A is secured: transport, schemes, auth and signing
Updated 18 November 2026 · Editorial assessment
What does A2A require for transport security?
HTTPS is mandatory in production, since all A2A communication in production environments must occur over HTTPS. The TLS version is a recommendation: implementations should use modern versions, and TLS 1.2 or higher is recommended.
Clients carry part of the load. They should verify the A2A server's identity by validating its TLS certificate against trusted certificate authorities during the handshake, and the page says this prevents man-in-the-middle attacks.
What security schemes can an agent card declare?
A card lists its authentication schemes in a security field that aligns with the OpenAPI Specification for authentication. Requirements sit in the card's securitySchemes and security fields, and the discovery page gives "Bearer" and "OAuth2" as example schemes. In version 1.0, the security scheme declarations were enhanced and gained mutual TLS support.
The changes page says the release added the Device Code flow and removed the deprecated implicit and password flows. A new pkce_required field on the Authorization Code flow indicates whether PKCE is mandatory.
How does a client authenticate to an A2A server?
A request moves through agent discovery, authentication, and the messaging APIs, in that order.
- In discovery, the client fetches the server's Agent Card to learn its capabilities and endpoint, and the standard location is /.well-known/agent-card.json on the server's domain.
- Next, it reads the Agent Card's security schemes and obtains a token when one is required.
- With a credential in hand, the caller sends messages and places the credential in standard HTTP headers as the chosen scheme requires.
Identity stays out of the message body. Payloads don't carry user or client identity information directly, and identity is established at the transport and HTTP layer.
Some tasks need a second credential mid-flight. When an agent needs access to another system during a task, the A2A server indicates to the client that more information is needed, and the client obtains those secondary credentials outside the protocol, for example through an OAuth flow.
What must the server authenticate and authorize?
Every request is authenticated, because the server must authenticate every incoming request using the credentials in its HTTP headers. Missing or invalid credentials get a 401 Unauthorized, which should include a WWW-Authenticate header. Valid credentials that lack permission for the action get a 403 Forbidden.
Authorization is also required on each operation. The specification says servers MUST implement authorization checks on every request and MUST scope results to the caller's authorized access boundaries. Those checks MUST occur before any database queries that could reveal resources outside the caller's scope.
The model itself is left to each agent. Authorization models may be based on user identity, roles, project membership or tenant boundaries. Access can be controlled on a per-skill basis, with OAuth scopes granting some skills and not others, and agents must grant only the necessary permissions.
For the identity design behind these checks, see the guide on agent identity and access.
How are push notifications secured?
Both ends share the work. The specification says agents (as webhook callers) and clients (as webhook receivers) have security responsibilities.
On the sending side, servers SHOULD NOT blindly trust and send POST requests to any URL provided by a client, since a malicious URL can lead to Server-Side Request Forgery. Listed mitigations are allowlisting of trusted domains, ownership verification, and network controls such as egress firewalls. The server MUST authenticate itself to the client's webhook URL using the scheme in TaskPushNotificationConfig.authentication, and common schemes include Bearer Tokens, API keys, HMAC signatures, or mTLS.
On the receiving side, the webhook MUST rigorously verify the authenticity of each incoming request. Notifications SHOULD include a timestamp so stale ones can be rejected, and for critical notifications the page suggests unique, single-use identifiers such as a JWT jti claim. In the documented example, the server makes its public keys available through a JWKS endpoint, and JWKS facilitates key rotation for asymmetric keys.
Can an agent card be signed?
Yes. Version 1.0 lists Agent Card signature verification using JWS and JSON Canonicalization among its enterprise features. The canonicalization standard, RFC 8785, provides deterministic JSON serialization for signing, and v1.0 uses JWS for the Agent Card's signatures field.
A card that contains sensitive data must be protected with authentication and authorization mechanisms, according to the discovery page. The discovery page recommends authenticated extended agent cards for sensitive information. It adds that the specification strongly recommends out-of-band dynamic credentials over static secrets embedded in a card.
What does A2A leave to the enterprise?
Registries, central policy, extension review and platform fit stay with you. Requirements writers can turn each item below into a line in the RFP checklist.
- Curated registries, which the discovery page describes for enterprise environments, require deploying and maintaining a registry service, and the current specification does not prescribe a standard API for curated registries.
- For servers exposed externally or across organizational boundaries, the enterprise page calls integration with API Management solutions highly recommended for authentication and authorization policy, rate limiting and quotas.
- Extensions widen the review scope, since new methods from an extension MUST be subject to the same authentication and authorization checks as core operations, and extension-related data from an external party is to be treated as untrusted input.
- Gemini Enterprise lets you register an A2A agent using either the Google Cloud console or the REST API. It doesn't use the securitySchemes and security fields of your card to authenticate and doesn't support the apiKey security scheme.
- Foundry Agent Service offers several A2A authentication methods, and its OAuth identity passthrough works with Microsoft and non-Microsoft A2A endpoints that support OAuth 2.0. Its docs also warn that anyone with access to the project can access secrets stored in a project connection.
- Amazon Bedrock AgentCore Runtime supports both SigV4 and OAuth 2.0 authentication schemes for A2A servers. Its troubleshooting section lists an authorization method mismatch and says to make sure a request uses the same method (OAuth or SigV4) that the agent was configured with.
Estates that span several of these runtimes can compare these differences with the patterns in connecting agents across clouds.
What do enterprise buyers ask?
Does mutual TLS work with A2A?
Mutual TLS support came with the enhanced security scheme declarations in v1.0. The discovery page also lists mTLS among the access controls on the HTTP endpoint serving the Agent Card.
Which OAuth flows were removed in v1.0?
The implicit and password flows were removed in v1.0. The changes page says the implicit flow was deprecated due to token leakage risks in browser history and logs.
Can one A2A endpoint serve several tenants?
Yes. Every request message has an optional tenant field whose value the server operator defines, and a gateway can also route on the authentication credentials already present in the request.
Where do I read the protocol background first?
Start with the A2A protocol overview, then return here for the security detail.
Which sources does this page cite?
- Enterprise Features (a2a-protocol.org)
- Overview (a2a-protocol.org)
- Multi-Tenancy (a2a-protocol.org)
- Agent Discovery (a2a-protocol.org)
- What's New in v1.0 (a2a-protocol.org)
- What is A2A? (a2a-protocol.org)
- Streaming & Asynchronous Operations (a2a-protocol.org)
- Overview (a2a-protocol.org)
- Register and manage A2A agents (docs.cloud.google.com)
- Agent2Agent (A2A) authentication (learn.microsoft.com)
- Deploy A2A servers in AgentCore Runtime (docs.aws.amazon.com)
Reviewed Nov 2026