top of page

TLS 1.3 Will Not Save the Internet, and TLS 1.4 Is Not Even Planned

Writer: Brian Couzens
Brian Couzens
2 minutes ago
15 min read
Brian Couzens and a colleague study a cracked holographic padlock glowing teal and red in a dark data centre aisle.

Post-quantum key exchange is arriving inside TLS 1.3, not behind a new protocol version. It protects new sessions against an adversary recording traffic today, on each terminated connection where both endpoints negotiate it. It does not protect traffic already harvested, the authentication of future sessions or the trust infrastructure the internet runs on, and a compliance check that looks only for TLS 1.3 cannot tell protected connections from classical ones.


Count the signatures in a typical publicly trusted browser handshake. Google's Chrome security team put a typical TLS connection at two public keys and five signatures to meet its authentication requirements. Every one of them is RSA or elliptic curve cryptography, including on connections that negotiate post-quantum key exchange.


That gap is the subject of this article. Transport Layer Security (TLS) 1.3 now carries hybrid post-quantum key exchange on about 70% of the browser-generated traffic reaching Cloudflare's network, by Cloudflare's own measure in September 2026, and that change matters. It protects one thing: the confidentiality of new sessions against an adversary who records traffic today, on each separately terminated TLS connection where both endpoints negotiate it. It does not protect traffic already recorded. It does not protect the authentication of future sessions. It does not protect the trust infrastructure the internet runs on, or the protocols that are not TLS. And no TLS 1.4 is chartered to draw a line between classical and post-quantum TLS.

SITG-Consulting has already set out how the hybrid handshake works in TLS 1.3. This piece sets out why deploying hybrid key exchange does not complete the internet's defence against harvest now, decrypt later (HNDL), or resolve the separate threat of harvest now, forge later (HNFL).


TLS 1.3 forward secrecy is not post-quantum secrecy

TLS 1.3 was a genuine improvement. RFC 9846, the July 2026 revision of the protocol, records that static RSA and Diffie-Hellman cipher suites were removed and that every public-key based key exchange mechanism now provides forward secrecy. It also encrypts every handshake message after the ServerHello.

Forward secrecy protects past sessions if a server's long-term private key is stolen later. It was designed against a thief, not against a quantum computer. In 1994 Peter Shor showed that a quantum computer could factor integers and compute discrete logarithms in polynomial time. The ephemeral elliptic curve key shares that give TLS 1.3 its forward secrecy travel in the ClientHello and ServerHello, which are sent unencrypted. For a full TLS 1.3 handshake using classical ephemeral key exchange without a pre-shared key, an adversary who records the handshake and the encrypted traffic can derive the traffic keys once a cryptographically relevant quantum computer (CRQC) exists.

Forward secrecy changes the attacker's arithmetic from one key per server to one computation per session. It does not remove the exposure. Published resource estimates have also fallen, although the requirements differ between factoring RSA and attacking elliptic curves. In May 2025 Google's Craig Gidney and Sophie Schmieg reported that RSA-2048 could in theory be factored by around a million noisy qubits running for about a week, roughly a twentieth of the physical qubits Google estimated in 2019.


There is no TLS 1.4, so there is no version boundary

The working group that maintains TLS has no new protocol version among its chartered work. Its charter sets three goals: extensions for emerging use cases, improvements to security, privacy and deployability, and maintenance of the current and previous versions. It also aims to minimise gratuitous changes to the protocol.

When the Internet Engineering Task Force (IETF) revised TLS 1.3 in July 2026, it kept the number. RFC 9846 describes itself as a minor update that retains the same version number and is backward compatible. SITG-Consulting covered that revision in TLS 1.3 hasn't become TLS 1.4. Post-quantum key exchange arrives as an optional extension inside TLS 1.3, standardised as RFC 10024 in August 2026. Post-quantum authentication for publicly trusted secure web (HTTPS) connections is also being developed within TLS 1.3.

Version numbers have given the industry a line to enforce before. In March 2021, RFC 8996 formally deprecated TLS 1.0 and 1.1 and moved both to Historic status, and browsers, regulators and auditors acted on the version. Standards can also tighten without a new number: RFC 9846 now states that TLS 1.0 and 1.1 MUST NOT be negotiated for any reason, while keeping the TLS 1.3 label. No equivalent requirement exists for post-quantum key exchange. RFC 10024 defines the hybrid groups; it is a definition, not a mandate to use them.

With no version boundary between classical and post-quantum TLS, a policy that demands "TLS 1.3 or higher" is satisfied by a connection running classical X25519. Classical negotiation remains valid wherever endpoint policy permits it. Endpoints can be configured to refuse classical groups and abort the handshake, but that is a configuration choice. Microsoft's own Windows, TLS 1.3 and Post Quantum Crypto FAQ describes the outcome where it is not made: if one side does not support hybrid post-quantum TLS, the connection may still succeed, but without hybrid post-quantum protection, with the exact behaviour depending on client, server and configuration. The endpoints know which group they negotiated. The user sees no difference, and neither does a compliance check that looks only for the version.


Post-quantum TLS 1.3 support is not end-to-end protection

Clients moved first. Chrome has offered the hybrid X25519MLKEM768 key exchange by default since Chrome 131. It pairs classical X25519 with ML-KEM, the post-quantum key encapsulation mechanism the National Institute of Standards and Technology (NIST) published as FIPS 203 in August 2024. Servers have not kept pace.

F5 Labs scanned the top one million domains for its 2026 State of PQC on the Web, published on 22 September 2026. It found that 54% of sites support post-quantum key exchange, falling to 22% once Cloudflare-hosted domains are excluded. These figures measure what websites support. They do not measure the share of traffic that negotiates post-quantum key exchange, or whether data is protected from end to end. A further 10.51% of responding sites do not support TLS 1.3 at all. Because RFC 10024 defines the hybrid groups for TLS 1.3 only, those sites have no standard route to post-quantum key exchange.

Behind the edge, the picture is thinner still. Cloudflare's scans of its customers' origin servers found that post-quantum support has grown from 0.5% in 2023 to 12.8%. Roughly seven in eight origins still cannot use post-quantum key exchange.


Each terminated connection stands alone

TLS protects the channel between two communicating peers, not the data's whole journey. Routers and transparent forwarding devices take no part in the key exchange. Every device that terminates TLS, whether a content delivery network edge, a load balancer, a reverse proxy, an inspection appliance or a service mesh, ends one connection and opens another with its own negotiation. A browser can negotiate X25519MLKEM768 with the edge while the edge negotiates classical X25519 with the origin. The first connection keeps its protection. The second is a separate exposure, and an adversary recording that leg records the same data under classical keys. Protection has to be assessed across every separately terminated connection that carries the data.

Inspection devices show the problem in miniature. Fortinet's own documentation records that, before its fix, hybrid post-quantum connections through FortiGate proxy-mode deep inspection were downgraded to non-ML-KEM encryption and the site would be unreachable. Support arrived in FortiOS 8.0 for proxy mode and in FortiOS 7.6.5 for flow mode. An inspection device that cannot negotiate the hybrid groups either breaks the connection or terminates it and negotiates classical cryptography on its own connections.

Each of those devices is a boundary where a new negotiation begins, and each is easy to miss in an inventory. SITG-Consulting's work on shadow cryptography and on interoperability in post-quantum cryptography sets out why.


Cutover does not reach back

Chrome only switched hybrid key exchange on by default for all desktop users as of Chrome 124, in 2024. Traffic recorded from connections using quantum-vulnerable key exchange remains exposed, and switching on ML-KEM today does nothing to protect those earlier recordings. Executive Order 14412, signed on 22 June 2026, names the risk directly: adversaries collecting United States information now, and decrypting it later.

Migration protects the future, not the archive an adversary already holds. Whether that archive matters depends on how long its contents must stay confidential. That is the question Mosca's theorem forces, and the reason SITG-Consulting argues that, for organisations treating this as a future problem, the data is already gone.


Key exchange defends against a recorder, not a forger

Google and Microsoft both explain why certificates are still classical. Google's Chrome team wrote in 2024 that there is no store-now-decrypt-later equivalent for authentication, so migrating authentication is less urgent than key exchange. Microsoft's FAQ calls classical certificates on post-quantum connections "by design", because an attacker would need a quantum computer at the time of the session. The IETF's own RFC 9954, published in July 2026, takes the same line: it treats post-quantum authentication separately from hybrid key exchange because session authentication cannot be broken retroactively.


That reasoning is correct for a recorded session, and it is the right basis for doing key exchange first. It is not a reason to treat authentication as later work, for three reasons.

First, the protection of future sessions depends on authentication. In TLS 1.3 the server signs a hash of the handshake transcript in its CertificateVerify message (RFC 9846, Section 4.5.2). The protocol's own security analysis states that the Finished message authentication code, together with that signature where present, prevents downgrade attacks on the negotiation. Once a CRQC exists, an attacker who recovers a server's private key from its public certificate, or a certificate authority's key from a published root, can impersonate that server in a new session wherever clients still accept the classical credential or trust path. The attacker completes the hybrid exchange with the client and reads the session as it happens. Sessions recorded earlier under hybrid key exchange stay protected. Hybrid key exchange protects recorded sessions against future decryption; it does not stop a future attacker impersonating a server whose authentication remains classical.

Second, the keys that matter longest are signing keys. Google's own researchers note that signature keys are harder to replace and longer lived, which makes them much more attractive targets when quantum computing time is scarce. Root certificates are public, and the public keys used to verify code and firmware signatures are distributed to the systems that rely on them. The private signing keys are not, but once a CRQC exists the public key is all an attacker needs to derive them.

Third, authentication takes longest to migrate, and it has to be finished before the machine arrives, not after. The next section shows how far the web is from that.

HNDL and HNFL are distinct attacks on the same classical public-key foundations, with different timelines, prerequisites and consequences. SITG-Consulting treats both as mandatory; its analysis of HNDL and HNFL sets out why long-lived data makes each viable. Both require a funded migration path, and neither outranks the other in general. In a given estate, sequencing follows data lifetime, trust dependencies, exposure and migration lead time, not an assumption that authentication can wait. The timing is not academic. One scenario-based preprint, published on arXiv by Grover, Haile, Pedersen, Uner and Erickson, models RSA risk crossing 50% between 2030 and 2032 under its stated assumptions. That is a conditional modelling result, not an established probability of a CRQC arriving in that window. SITG-Consulting has also set out how Q-Day would actually unfold: not as a breach that can be patched.


The web cannot migrate one organisation at a time

Two parties can upgrade a key exchange between themselves. Authentication on the public web depends on certificate authorities, browser root programmes, Certificate Transparency logs and the CA/Browser Forum, where certificate authorities and browser makers agree the rules for publicly trusted certificates. None of them has finished.

Three things have to move, and they move separately: the certificate chain the client validates, the server's CertificateVerify signature proving it holds the private key, and the client policy that decides whether classical alternatives are still accepted. Talk of "post-quantum certificates" alone hides the third.

Chrome announced in February 2026 that it has no immediate plan to add traditional X.509 certificates containing post-quantum cryptography to the Chrome Root Store. It is backing Merkle Tree Certificates instead, which replace the chain of signatures with a compact proof that a certificate is included in a signed tree. Its plan invites Certificate Transparency log operators to bootstrap public Merkle Tree Certificates in its second phase, from the first quarter of 2027.

Size is the reason. Post-quantum signatures are, in Cloudflare's words, roughly 40 times larger than classical ones, and Cloudflare estimates they would multiply the data Certificate Transparency logs must store by the same factor. F5 Labs found no post-quantum certificates on the public web, and estimates a certificate chain signed with ML-DSA-65, a parameter set of NIST's post-quantum signature standard FIPS 204, at 12.2 KB to 18.3 KB, against about 2 KB to 3 KB for the Elliptic Curve Digital Signature Algorithm (ECDSA). Cloudflare's own Merkle Tree Certificate authority is targeting inclusion in Chrome's quantum-resistant root store in early 2027, subject to evaluation, and the company says it plans to upgrade to post-quantum cryptography by 2029.

The policy body is slower still. The CA/Browser Forum minuted concerns about the scalability of the Certificate Transparency ecosystem during its discussion of an ML-DSA ballot in May 2026. In July its proposers were still reconciling two competing drafts. On 22 October the Forum's Server Certificate Working Group recorded that draft ballot SC-106, "Allow ML-DSA", would be removed from its tracking list, with discussion potentially moving to the forthcoming Client Authentication Working Group.

The structural point is the one Cloudflare makes itself. A domain that moves to post-quantum authentication should still monitor the logs for legacy certificates issued in its name, because clients that accept them leave a malicious downgrade path open. Monitoring is a detection control, not a prevention control: it finds a certificate that was logged, it does not stop a client accepting one, and the Certificate Transparency ecosystem itself still relies on elliptic curve P-256 signatures. One organisation's post-quantum certificate does not protect its name while browsers still trust classical certificates for it. That is an ecosystem problem, and it lands on public key infrastructure (PKI) already being reshaped by the 47-day certificate schedule. The Merkle tree itself is not new; SITG-Consulting has explained why it matters more today than ever.


The internet is not one protocol

TLS carries the web. It does not carry the internet's routing, naming or software supply chain.

Start with routing. The Resource Public Key Infrastructure (RPKI), which lets networks verify who may originate a block of IP addresses, mandates RSA signatures with a 2048-bit modulus. Post-quantum work for RPKI exists only in an individual Internet-Draft, which does not update that profile or authorise production use.

Naming is no further ahead. The DNS Security Extensions (DNSSEC) sign with classical algorithms such as RSA, ECDSA and the Edwards-curve Digital Signature Algorithm (EdDSA). The IANA DNSSEC algorithm registry lists a single post-quantum entry, ML-DSA-44, and its reference is an informational Internet-Draft.

Privacy protection inherits the same gap. Encrypted Client Hello (RFC 9849, March 2026) hides the name of the site a user visits by encrypting the ClientHello under the server's public key, using Hybrid Public Key Encryption (HPKE). The HPKE standard, RFC 9180, defines only elliptic curve Diffie-Hellman key encapsulation mechanisms. Post-quantum ones for HPKE are still an Internet-Draft. A server name hidden under a classical key today is open to harvest now, decrypt later.

Software and firmware signing rest on the same foundation. Code-signing, firmware-signing and update-signing keys that use RSA, ECDSA or EdDSA fall under the same timetable: NIST deprecates those algorithms after 2030 and disallows them after 2035 in its final IR 8547. Verifying signatures that already exist remains permitted as legacy use after that point; creating new ones with those algorithms does not.

Not all of it is lagging. OpenSSH has offered post-quantum key agreement by default since version 9.0 in April 2022, made hybrid ML-KEM its default in version 10.0 in April 2025, and from version 10.1 warns users when a connection does not use post-quantum key exchange. There is no comparable warning in the browser's address bar. To see which key exchange a page used in Chrome, Cloudflare's instructions are to right-click, choose Inspect and open the Security tab.


Governance: the padlock says nothing

The protocol will not tell a board whether its estate is protected, and neither will the browser. Assurance has to come from the estate itself. Until September 2026, Cloudflare had not exposed domain-level information on the algorithms its customers' traffic used, so those customers could not answer whether their domains were post-quantum encrypted.

That changes what governance must ask for.

Policies and audits must name algorithms, not versions. "TLS 1.3 enabled" is not a post-quantum control. The control is the negotiated key exchange group and signature algorithm, recorded per service and per terminated connection.

Evidence must cover every termination point, inside the network as well as at the edge. A test from the internet describes the edge; it says nothing about the service behind it. A Cryptographic Bill of Materials (CBOM) should record what each point negotiates, not what it is configured to offer, which is the discipline set out in The Event Horizon II. Deploying PQC is not the same as proving it works.

Contracts must carry the requirement. Content delivery network to origin connections, inspection services, managed firewalls and certificate authorities all sit outside the organisation's direct control. SITG-Consulting's guidance on cryptographic debt in procurement applies here.

Ownership sits with risk. SITG-Consulting's recommended accountability model treats post-quantum transformation as an enterprise risk programme: executive accountability with the Chief Risk Officer, delivery with the Chief Information Security Officer and engineering. SITG-Consulting has argued why governance comes before discovery.

The dates are on the record. NIST's final IR 8547, published on 27 October 2026, deprecates quantum-vulnerable signature and key-establishment algorithms after 2030 and disallows their use under NIST standards and guidelines after 2035. It is written for federal agencies, technology providers, standards organisations and testing laboratories; it is not a universal legal deadline. SITG-Consulting's digest of the final report covers the change. For United States federal high value assets and high impact systems, Executive Order 14412 requires post-quantum key establishment by 31 December 2030 and post-quantum signatures by 31 December 2031, excluding National Security Systems.


What to do now

  1. Rewrite TLS policy in algorithms, not versions. Require a post-quantum hybrid group from RFC 10024 on named services, and record every connection that still negotiates a classical group as an exception with an owner and an expiry date.

  2. Map every point where TLS terminates: edge, origin, load balancers, reverse proxies, inspection devices, virtual private network (VPN) gateways and service meshes. Record the key exchange group each one negotiates, tested from inside the network and from outside.

  3. Retire TLS 1.2 wherever it can be retired. It has no standard route to post-quantum key exchange.

  4. Put post-quantum support for origin connections, inspection and managed services into contracts and renewal criteria.

  5. Start the authentication programme now. Inventory public and private PKI, certificate lifetimes, hardware security module support for ML-DSA, the dependency on Certificate Transparency and the client policies that still accept classical alternatives. Track Chrome's Merkle Tree Certificate programme and changes to the CA/Browser Forum's post-quantum certificate requirements.

  6. Inventory long-lived signing keys, including code, firmware, document and update signing, with a replacement path for each.

  7. Extend the scope beyond TLS: Secure Shell (SSH) configuration, DNSSEC signing, RPKI exposure through network providers, Encrypted Client Hello keys and VPNs.

  8. Assess what has already been exposed. Identify data whose required confidentiality outlives the likely arrival of a CRQC and which crossed classical links.

  9. Keep evidence. One handshake that negotiates X25519MLKEM768 is a test result, not proof across an estate.

  10. Give the programme a risk owner, and report it to the board as a risk, not as a technology upgrade.


The evidence test

TLS 1.3 did what it was built to do. It was not built to carry the internet through the arrival of a CRQC, and there is no TLS 1.4 waiting to finish the job. Hybrid key exchange closes one door, on one connection, against one threat. Completing the defence takes post-quantum authentication, infrastructure migration and governance that demands evidence.

An organisation that believes switching on hybrid key exchange has settled its exposure should be able to produce the negotiated groups, the authentication paths and the exceptions across its estate. "TLS 1.3 enabled" is not that evidence.


About the author

Brian Couzens is CEO of SITG-Consulting, an independent advisory firm working on post-quantum transformation, cryptographic governance and assurance. SITG-Consulting provides cryptographic agility services and independent PQC assurance.


Sources

Comments


Need an independent assessment of your cryptographic exposure?

bottom of page