RFC 9325: BCP 195: Recommendations for Secure Use of Transport Layer Security (TLS) and Datagram Transport Layer Security (DTLS)

3.1. Protocol Versions

3.2. Strict TLS

The following recommendations are provided to help prevent "SSL Stripping" and STARTTLS command injection (attacks that are summarized in [RFC7457]):¶

Rationale: Combining unprotected and TLS-protected communication opens the way to SSL Stripping and similar attacks, since an initial part of the communication is not integrity protected and therefore can be manipulated by an attacker whose goal is to keep the communication in the clear.¶

3.3. Compression

Rationale: TLS compression has been subject to security attacks such as the Compression Ratio Info-leak Made Easy (CRIME) attack.¶

Implementers should note that compression at higher protocol levels can allow an active attacker to extract cleartext information from the connection. The Browser Reconnaissance and Exfiltration via Adaptive Compression of Hypertext (BREACH) attack is one such case. These issues can only be mitigated outside of TLS and are thus outside the scope of this document. See Section 2.6 of [RFC7457] for further details.¶

3.4. TLS Session Resumption

Session resumption drastically reduces the number of full TLS handshakes and thus is an essential performance feature for most deployments.¶

Stateless session resumption with session tickets is a popular strategy. For TLS 1.2, it is specified in [RFC5077]. For TLS 1.3, a more secure mechanism based on the use of a pre-shared key (PSK) is described in Section 4.6.1 of [RFC8446]. See [Springall16] for a quantitative study of the risks induced by TLS cryptographic "shortcuts", including session resumption.¶

When it is used, the resumption information MUST be authenticated and encrypted to prevent modification or eavesdropping by an attacker. Further recommendations apply to session tickets:¶

Rationale: Session resumption is another kind of TLS handshake and therefore must be as secure as the initial handshake. This document (Section 4) recommends the use of cipher suites that provide forward secrecy, i.e., that prevent an attacker who gains momentary access to the TLS endpoint (either client or server) and its secrets from reading either past or future communication. The tickets must be managed so as not to negate this security property.¶

TLS 1.3 provides the powerful option of forward secrecy even within a long-lived connection that is periodically resumed. Section 2.2 of [RFC8446] recommends that clients SHOULD send a "key_share" when initiating session resumption. In order to gain forward secrecy, this document recommends that server implementations SHOULD select the "psk_dhe_ke" PSK key exchange mode and respond with a "key_share" to complete an Ephemeral Elliptic Curve Diffie-Hellman (ECDHE) exchange on each session resumption. As a more performant alternative, server implementations MAY refrain from responding with a "key_share" until a certain amount of time (e.g., measured in hours) has passed since the last ECDHE exchange; this implies that the "key_share" operation would not occur for the presumed majority of session resumption requests (which would occur within a few hours) while still ensuring forward secrecy for longer-lived sessions.¶

TLS session resumption introduces potential privacy issues where the server is able to track the client, in some cases indefinitely. See [Sy2018] for more details.¶

3.5. Renegotiation in TLS 1.2

The recommendations in this section apply to TLS 1.2 only, because renegotiation has been removed from TLS 1.3.¶

Renegotiation in TLS 1.2 is a handshake that establishes new cryptographic parameters for an existing session. The mechanism existed in TLS 1.2 and in earlier protocol versions and was improved following several major attacks including a plaintext injection attack, CVE-2009-3555 [CVE].¶

TLS 1.2 clients and servers MUST implement the renegotiation_info extension, as defined in [RFC5746].¶

TLS 1.2 clients MUST send renegotiation_info in the Client Hello. If the server does not acknowledge the extension, the client MUST generate a fatal handshake_failure alert prior to terminating the connection.¶

Rationale: It is not safe for a client to connect to a TLS 1.2 server that does not support renegotiation_info regardless of whether either endpoint actually implements renegotiation. See also Section 4.1 of [RFC5746].¶

A related attack resulting from TLS session parameters not being properly authenticated is a Triple Handshake [Triple-Handshake]. To address this attack, TLS 1.2 implementations MUST support the extended_master_secret extension defined in [RFC7627].¶

3.6. Post-Handshake Authentication

Renegotiation in TLS 1.2 was (partially) replaced in TLS 1.3 by separate post-handshake authentication and key update mechanisms. In the context of protocols that multiplex requests over a single connection (such as HTTP/2 [RFC9113]), post-handshake authentication has the same problems as TLS 1.2 renegotiation. Multiplexed protocols SHOULD follow the advice provided for HTTP/2 in Section 9.2.3 of [RFC9113].¶

3.7. Server Name Indication (SNI)

TLS implementations MUST support the Server Name Indication (SNI) extension defined in Section 3 of [RFC6066] for those higher-level protocols that would benefit from it, including HTTPS. However, the actual use of SNI in particular circumstances is a matter of local policy. At the time of writing, a technology for encrypting the SNI (called Encrypted Client Hello) is being worked on in the TLS Working Group [TLS-ECH]. Once that method has been standardized and widely implemented, it will likely be appropriate to recommend its usage in a future version of this BCP.¶

Rationale: SNI supports deployment of multiple TLS-protected virtual servers on a single address, and therefore enables fine-grained security for these virtual servers, by allowing each one to have its own certificate. However, SNI also leaks the target domain for a given connection; this information leak will be closed by use of TLS Encrypted Client Hello once that method has been standardized.¶

In order to prevent the attacks described in [ALPACA], a server that does not recognize the presented server name SHOULD NOT continue the handshake and instead SHOULD fail with a fatal-level unrecognized_name(112) alert. Note that this recommendation updates Section 3 of [RFC6066], which stated:¶

If the server understood the ClientHello extension but does not recognize the server name, the server SHOULD take one of two actions: either abort the handshake by sending a fatal-level unrecognized_name(112) alert or continue the handshake.¶

Clients SHOULD abort the handshake if the server acknowledges the SNI extension but presents a certificate with a different hostname than the one sent by the client.¶

3.8. Application-Layer Protocol Negotiation (ALPN)

TLS implementations (both client- and server-side) MUST support the Application-Layer Protocol Negotiation (ALPN) extension [RFC7301].¶

In order to prevent "cross-protocol" attacks resulting from failure to ensure that a message intended for use in one protocol cannot be mistaken for a message for use in another protocol, servers are advised to strictly enforce the behavior prescribed in Section 3.2 of [RFC7301]:¶

In the event that the server supports no protocols that the client advertises, then the server SHALL respond with a fatal 'no_application_protocol' alert.¶

Clients SHOULD abort the handshake if the server acknowledges the ALPN extension but does not select a protocol from the client list. Failure to do so can result in attacks such those described in [ALPACA].¶

Protocol developers are strongly encouraged to register an ALPN identifier for their protocols. This applies both to new protocols and to well-established protocols; however, because the latter might have a large deployed base, strict enforcement of ALPN usage may not be feasible when an ALPN identifier is registered for a well-established protocol.¶

3.9. Multi-Server Deployment

Deployments that involve multiple servers or services can increase the size of the attack surface for TLS. Two scenarios are of interest:¶

3.10. Zero Round-Trip Time (0-RTT) Data in TLS 1.3

The 0-RTT early data feature is new in TLS 1.3. It provides reduced latency when TLS connections are resumed, at the potential cost of certain security properties. As a result, it requires special attention from implementers on both the server and the client side. Typically, this extends to the TLS library as well as protocol layers above it.¶

For HTTP over TLS, refer to [RFC8470] for guidance.¶

For QUIC on TLS, refer to Section 9.2 of [RFC9001].¶

For other protocols, generic guidance is given in Section 8 and Appendix E.5 of [RFC8446]. To paraphrase Appendix E.5, applications MUST avoid this feature unless an explicit specification exists for the application protocol in question to clarify when 0-RTT is appropriate and secure. This can take the form of an IETF RFC, a non-IETF standard, or documentation associated with a non-standard protocol.¶

4.1. General Guidelines

Cryptographic algorithms weaken over time as cryptanalysis improves: algorithms that were once considered strong become weak. Consequently, cipher suites using weak algorithms need to be phased out and replaced with more secure cipher suites. This helps to ensure that the desired security properties still hold. SSL/TLS has been in existence for well over 20 years and many of the cipher suites that have been recommended in various versions of SSL/TLS are now considered weak or at least not as strong as desired. Therefore, this section modernizes the recommendations concerning cipher suite selection.¶

4.2. Cipher Suites for TLS 1.2

Given the foregoing considerations, implementation and deployment of the following cipher suites is RECOMMENDED:¶

As these are Authenticated Encryption with Associated Data (AEAD) algorithms [RFC5116], these cipher suites are supported only in TLS 1.2 and not in earlier protocol versions.¶

Typically, to prefer these suites, the order of suites needs to be explicitly configured in server software. It would be ideal if server software implementations were to prefer these suites by default.¶

Some devices have hardware support for AES Counter Mode with CBC-MAC (AES-CCM) but not AES Galois/Counter Mode (AES-GCM), so they are unable to follow the foregoing recommendations regarding cipher suites. There are even devices that do not support public key cryptography at all, but these are out of scope entirely.¶

A cipher suite that operates in CBC (cipher block chaining) mode (e.g., TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256) SHOULD NOT be used unless the encrypt_then_mac extension [RFC7366] is also successfully negotiated. This requirement applies to both client and server implementations.¶

When using ECDSA signatures for authentication of TLS peers, it is RECOMMENDED that implementations use the NIST curve P-256. In addition, to avoid predictable or repeated nonces (which could reveal the long-term signing key), it is RECOMMENDED that implementations implement "deterministic ECDSA" as specified in [RFC6979] and in line with the recommendations in [RFC8446].¶

Note that implementations of "deterministic ECDSA" may be vulnerable to certain side-channel and fault injection attacks precisely because of their determinism. While most fault injection attacks described in the literature assume physical access to the device (and therefore are more relevant in Internet of Things (IoT) deployments with poor or non-existent physical security), some can be carried out remotely [Poddebniak2017], e.g., as Rowhammer [Kim2014] variants. In deployments where side-channel attacks and fault injection attacks are a concern, implementation strategies combining both randomness and determinism (for example, as described in [CFRG-DET-SIGS]) can be used to avoid the risk of successful extraction of the signing key.¶

4.3. Cipher Suites for TLS 1.3

This document does not specify any cipher suites for TLS 1.3. Readers are referred to Section 9.1 of [RFC8446] for cipher suite recommendations.¶

4.4. Limits on Key Usage

All ciphers have an upper limit on the amount of traffic that can be securely protected with any given key. In the case of AEAD cipher suites, two separate limits are maintained for each key:¶

The latter applies to DTLS (and also to QUIC) but not to TLS itself, since TLS connections are torn down on the first decryption failure.¶

When a sender is approaching CL, the implementation SHOULD initiate a new handshake (in TLS 1.3, this can be achieved by sending a KeyUpdate message on the established session) to rotate the session key. When a receiver has reached IL, the implementation SHOULD close the connection. Although these recommendations are a best practice, implementers need to be aware that it is not always easy to accomplish them in protocols that are built on top of TLS/DTLS without introducing coordination across layer boundaries. See Section 6 of [RFC9001] for an example of the cooperation that was necessary in QUIC between the crypto and transport layers to support key updates. Note that in general, application protocols might not be able to emulate that method given their more constrained interaction with TLS/DTLS. As a result of these complexities, these recommendations are not mandatory.¶

For all TLS 1.3 cipher suites, readers are referred to Section 5.5 of [RFC8446] for the values of CL and IL. For all DTLS 1.3 cipher suites, readers are referred to Section 4.5.3 of [RFC9147].¶

For all AES-GCM cipher suites recommended for TLS 1.2 and DTLS 1.2 in this document, CL can be derived by plugging the corresponding parameters into the inequalities in Section 6.1 of [AEAD-LIMITS] that apply to random, partially implicit nonces, i.e., the nonce construction used in TLS 1.2. Although the obtained figures are slightly higher than those for TLS 1.3, it is RECOMMENDED that the same limit of 2^24.5 records is used for both versions.¶

For all AES-GCM cipher suites recommended for DTLS 1.2, IL (obtained from the same inequalities referenced above) is 2²⁸.¶

4.5. Public Key Length

When using the cipher suites recommended in this document, two public keys are normally used in the TLS handshake: one for the Diffie-Hellman key agreement and one for server authentication. Where a client certificate is used, a third public key is added.¶

With a key exchange based on modular exponential (MODP) Diffie-Hellman groups ("DHE" cipher suites), DH key lengths of at least 2048 bits are REQUIRED.¶

Rationale: For various reasons, in practice, DH keys are typically generated in lengths that are powers of two (e.g., 2¹⁰ = 1024 bits, 2¹¹ = 2048 bits, 2¹² = 4096 bits). Because a DH key of 1228 bits would be roughly equivalent to only an 80-bit symmetric key [RFC3766], it is better to use keys longer than that for the "DHE" family of cipher suites. A DH key of 1926 bits would be roughly equivalent to a 100-bit symmetric key [RFC3766]. A DH key of 2048 bits (equivalent to a 112-bit symmetric key) is the minimum allowed by the latest revision of [NIST.SP.800-56A] as of this writing (see in particular Appendix D of that document).¶

As noted in [RFC3766], correcting for the emergence of The Weizmann Institute Relation Locator (TWIRL) machine [TWIRL] would imply that 1024-bit DH keys yield about 61 bits of equivalent strength and that a 2048-bit DH key would yield about 92 bits of equivalent strength. The Logjam attack [Logjam] further demonstrates that 1024-bit Diffie-Hellman parameters should be avoided.¶

With regard to ECDH keys, implementers are referred to the IANA "TLS Supported Groups" registry (formerly known as the "EC Named Curve Registry") within the "Transport Layer Security (TLS) Parameters" registry [IANA_TLS] and in particular to the "recommended" groups. Curves of less than 224 bits MUST NOT be used. This recommendation is in line with the latest revision of [NIST.SP.800-56A].¶

When using RSA, servers MUST authenticate using certificates with at least a 2048-bit modulus for the public key. In addition, the use of the SHA-256 hash algorithm is RECOMMENDED and SHA-1 or MD5 MUST NOT be used [RFC9155] (for more details, see also [CAB-Baseline], for which the current version at the time of writing is 1.8.4). Clients MUST indicate to servers that they request SHA-256 by using the "Signature Algorithms" extension defined in TLS 1.2. For TLS 1.3, the same requirement is already specified by [RFC8446].¶

4.6. Truncated HMAC

Implementations MUST NOT use the Truncated HMAC Extension, defined in Section 7 of [RFC6066].¶

Rationale: The extension does not apply to the AEAD cipher suites recommended above. However, it does apply to most other TLS cipher suites. Its use has been shown to be insecure in [PatersonRS11].¶

5.1. Security Services

This document provides recommendations for an audience that wishes to secure their communication with TLS to achieve the following:¶

Confidentiality:

all application-layer communication is encrypted with the goal that no party should be able to decrypt it except the intended receiver.¶

Data integrity:

any changes made to the communication in transit are detectable by the receiver.¶

Authentication:

an endpoint of the TLS communication is authenticated as the intended entity to communicate with.¶

With regard to authentication, TLS enables authentication of one or both endpoints in the communication. In the context of opportunistic security [RFC7435], TLS is sometimes used without authentication. As discussed in Section 5.2, considerations for opportunistic security are not in scope for this document.¶

If deployers deviate from the recommendations given in this document, they need to be aware that they might lose access to one of the foregoing security services.¶

This document applies only to environments where confidentiality is required. It requires algorithms and configuration options that enforce secrecy of the data in transit.¶

This document also assumes that data integrity protection is always one of the goals of a deployment. In cases where integrity is not required, it does not make sense to employ TLS in the first place. There are attacks against confidentiality-only protection that utilize the lack of integrity to also break confidentiality (see, for instance, [DegabrieleP07] in the context of IPsec).¶

This document addresses itself to application protocols that are most commonly used on the Internet with TLS and DTLS. Typically, all communication between TLS clients and TLS servers requires all three of the above security services. This is particularly true where TLS clients are user agents like web browsers or email clients.¶

This document does not address the rarer deployment scenarios where one of the above three properties is not desired, such as the use case described in Section 5.2. As another scenario where confidentiality is not needed, consider a monitored network where the authorities in charge of the respective traffic domain require full access to unencrypted (plaintext) traffic and where users collaborate and send their traffic in the clear.¶

5.2. Opportunistic Security

There are several important scenarios in which the use of TLS is optional, i.e., the client decides dynamically ("opportunistically") whether to use TLS with a particular server or to connect in the clear. This practice, often called "opportunistic security", is described at length in [RFC7435] and is often motivated by a desire for backward compatibility with legacy deployments.¶

In these scenarios, some of the recommendations in this document might be too strict, since adhering to them could cause fallback to cleartext, a worse outcome than using TLS with an outdated protocol version or cipher suite.¶

7.1. Host Name Validation

Application authors should take note that some TLS implementations do not validate host names. If the TLS implementation they are using does not validate host names, authors might need to write their own validation code or consider using a different TLS implementation.¶

It is noted that the requirements regarding host name validation (and, in general, binding between the TLS layer and the protocol that runs above it) vary between different protocols. For HTTPS, these requirements are defined by Sections 4.3.3, 4.3.4, and 4.3.5 of [RFC9110].¶

Host name validation is security-critical for all common TLS use cases. Without it, TLS ensures that the certificate is valid and guarantees possession of the private key but does not ensure that the connection terminates at the desired endpoint. Readers are referred to [RFC6125] for further details regarding generic host name validation in the TLS context. In addition, that RFC contains a long list of application protocols, some of which implement a policy very different from HTTPS.¶

If the host name is discovered indirectly and insecurely (e.g., by a cleartext DNS query for an SRV or Mail Exchange (MX) record), it SHOULD NOT be used as a reference identifier [RFC6125] even when it matches the presented certificate. This proviso does not apply if the host name is discovered securely (for further discussion, see [RFC7673] and [RFC7672]).¶

Host name validation typically applies only to the leaf "end entity" certificate. Naturally, in order to ensure proper authentication in the context of the PKI, application clients need to verify the entire certification path in accordance with [RFC5280].¶

7.2. AES-GCM

Section 4.2 recommends the use of the AES-GCM authenticated encryption algorithm. Please refer to Section 6 of [RFC5288] for security considerations that apply specifically to AES-GCM when used with TLS.¶

7.3. Forward Secrecy

Forward secrecy (also called "perfect forward secrecy" or "PFS" and defined in [RFC4949]) is a defense against an attacker who records encrypted conversations where the session keys are only encrypted with the communicating parties' long-term keys.¶

Should the attacker be able to obtain these long-term keys at some point later in time, the session keys and thus the entire conversation could be decrypted.¶

In the context of TLS and DTLS, such compromise of long-term keys is not entirely implausible. It can happen, for example, due to:¶

Forward secrecy ensures in such cases that it is not feasible for an attacker to determine the session keys even if the attacker has obtained the long-term keys some time after the conversation. It also protects against an attacker who is in possession of the long-term keys but remains passive during the conversation.¶

Forward secrecy is generally achieved by using the Diffie-Hellman scheme to derive session keys. The Diffie-Hellman scheme has both parties maintain private secrets and send parameters over the network as modular powers over certain cyclic groups. The properties of the so-called Discrete Logarithm Problem (DLP) allow the parties to derive the session keys without an eavesdropper being able to do so. There is currently no known attack against DLP if sufficiently large parameters are chosen. A variant of the Diffie-Hellman scheme uses elliptic curves instead of the originally proposed modular arithmetic. Given the current state of the art, Elliptic Curve Diffie-Hellman appears to be more efficient, permits shorter key lengths, and allows less freedom for implementation errors than finite-field Diffie-Hellman.¶

Unfortunately, many TLS/DTLS cipher suites were defined that do not feature forward secrecy, e.g., TLS_RSA_WITH_AES_256_CBC_SHA256. This document therefore advocates strict use of forward-secrecy-only ciphers.¶

7.4. Diffie-Hellman Exponent Reuse

For performance reasons, it is not uncommon for TLS implementations to reuse Diffie-Hellman and Elliptic Curve Diffie-Hellman exponents across multiple connections. Such reuse can result in major security issues:¶

To address these concerns:¶

7.5. Certificate Revocation

The following considerations and recommendations represent the current state of the art regarding certificate revocation, even though no complete and efficient solution exists for the problem of checking the revocation status of common public key certificates [RFC5280]:¶

For the common use cases of public key certificates in TLS, servers SHOULD support the following as a best practice given the current state of the art and as a foundation for a possible future solution: OCSP [RFC6960] and OCSP stapling using the status_request extension defined in [RFC6066]. Note that the exact mechanism for embedding the status_request extension differs between TLS 1.2 and 1.3. As a matter of local policy, server operators MAY request that CAs issue must-staple [RFC7633] certificates for the server and/or for client authentication, but we recommend reviewing the operational conditions before deciding on this approach.¶

The considerations in this section do not apply to scenarios where the DNS-Based Authentication of Named Entities (DANE) TLSA resource record [RFC6698] is used to signal to a client which certificate a server considers valid and good to use for TLS connections.¶