Definition of mTLS, or mutual TLS authentication
In an ordinary TLS connection, such as that of a browser to an HTTPS website, only the server proves its identity with a certificate. mTLS, for mutual TLS, adds the reverse proof: the client also presents an X.509 certificate and demonstrates that it holds the corresponding private key. IETF RFC 8705 (February 2020) defines the expression in this way in its section 1.2.
The mechanism belongs to TLS. The TLS 1.3 specification, RFC 9846 (July 2026, which replaces RFC 8446), states that the server is always authenticated and the client optionally (section 1). The acronym mTLS is a matter of usage. These RFCs were read on rfc-editor.org on 12 November 2026.
Three more messages in the handshake
- With the CertificateRequest message, the server requests a certificate from the client (RFC 9846, section 4.4.2).
- The client responds with the Certificate message, which carries its certificate chain (section 4.5.1).
- It adds CertificateVerify, a signature that proves possession of the private key (section 4.5.2).
If the client presents no certificate, the server chooses: continue without authenticating it, or abort the handshake with the “certificate_required” alert (section 4.5.1.3). Requiring mTLS means choosing the second option.
What mTLS is for: authenticating machines to each other
mTLS is used where no person types in a password: calls between microservices, exchanges between partners, connection of a device. The identity is that of the certificate, issued by a certificate authority that the other end recognises.
Service meshes such as Istio or Linkerd establish it between their proxies. According to their documentation, consulted on the same day, a built-in certificate authority issues a certificate to each proxy and renews it automatically. Both also describe a regime in which plaintext traffic remains accepted: the so-called permissive mode in Istio, traffic from sources outside the mesh in Linkerd. Installing a mesh therefore does not mean mTLS everywhere.
Take two services of an online retailer: the order service calls the payment service. Once mTLS is required, the payment service refuses any connection without a recognised certificate, for example from a container introduced into the same network. Admitting only the identity of the order service is already a matter of an authorisation rule, as is the case where that service, itself compromised, presents a certificate that is still valid.
mTLS, plain TLS, VPN, access token: not to be confused
| Mechanism | What is authenticated | Scope | What remains to be handled |
|---|---|---|---|
| Plain TLS | The server | One connection | The client’s identity |
| mTLS | The server and the client | One connection | The client’s rights, the certificate life cycle |
| VPN | The workstation or site joining a network | A tunnel to that network | Access to each service on that network |
| Access token | The token holder, whoever they are | One request | Token theft |
mTLS authenticates without authorising: knowing who is calling does not say what they are entitled to do. On its own, it does not constitute a zero trust architecture. It can be combined with tokens: RFC 8705 binds an OAuth access token to the client certificate.
The trade-off: certificates to keep alive
Each service holds a private key and a certificate, which must be issued, distributed, renewed before expiry and revoked in the event of compromise. The authority’s own certificate must also be renewed. An expired certificate can lead to connections being refused, and therefore to an outage.
Another limit: when TLS terminates on a load balancer or reverse proxy, RFC 8705 leaves out of its scope how the client’s identity is passed on to the application (section 6.5).
Automating this life cycle falls to the platform team, in the sense of our DevOps and platform engineering page. The certificate authority becomes a dependency of every exchange: its rank in the recovery order after an incident is set with the plans on our cyber resilience page.