How much post-quantum key agreement adds to a TLS handshake
In this piece
- Hybrid key agreement is standardised but post-quantum TLS signatures are not
- Each key share grows from 32 bytes to more than a thousand
- Hybrids do not protect against a shared random-number generator
- Policy dates and quantum roadmaps are not evidence of deployment
- Every source is tagged by the kind of evidence it is
- Still open
Adding post-quantum key agreement to TLS, the protocol that secures web connections, makes the client's key share (the public value it sends so both sides can agree a secret) 1,216 bytes where a classical X25519 share is 32. That comes from the first complete review of the post-quantum landscape in this research, written at 01:41 on 14 October. It describes itself as a targeted narrative review of primary sources rather than a systematic one, with no experiments and no replication.
An earlier review, at 00:40 the same day, was never finished. Its full text could not be saved because GitHub writes were blocked, and the run log marks the placeholder as unfinished.
Hybrid key agreement is standardised but post-quantum TLS signatures are not
RFC 10024 (August 2026, Standards Track) defines hybrid key agreement for TLS 1.3, which pairs a classical algorithm with ML-KEM, a post-quantum key-encapsulation mechanism, as three named groups: X25519MLKEM768, SecP256r1MLKEM768 and SecP384r1MLKEM1024. RFC 9954 (July 2026) is the generic construction, and it is Informational. NIST SP 800-227, final guidance on using key-encapsulation mechanisms, is dated 18 September 2025. HQC was selected by NIST in March 2025, but the review found no primary evidence that it had become a final FIPS standard, and says so rather than assuming.
Signatures are a separate migration. RFC 9881 (October 2025) gives identifiers for ML-DSA, a post-quantum signature scheme, in X.509 certificates. The TLS document for ML-DSA, dated 11 September 2026, is still an Internet-Draft.
Each key share grows from 32 bytes to more than a thousand
A classical X25519 key share is 32 bytes. The hybrid client share is 1,216 bytes (a 1,184-byte ML-KEM-768 encapsulation key plus 32 bytes of X25519) and the server share is 1,120 bytes (a 1,088-byte ciphertext plus 32). The combined input secret is 64 bytes.
The review ran no latency benchmark, and lists one as an open experiment.
The order of the parts differs: for X25519MLKEM768 the ML-KEM part is serialised first, for the two NIST-curve hybrids the elliptic-curve part comes first. Implementations must validate the ML-KEM key, check ciphertext lengths and elliptic-curve points, and reject an all-zero X25519 secret. The review classes these as correctness conditions rather than optimisations.
Hybrids do not protect against a shared random-number generator
The RFC warns that if one insecure random-number generator feeds both halves of a hybrid, a compromise can hit both at once. Hybridisation protects against one algorithm failing, but a flaw shared by both halves defeats it. The review turns that into a research question: model the shared-generator case in a formal TLS 1.3 security game (a precise model of what an attacker can do) and find which claims survive.
Policy dates and quantum roadmaps are not evidence of deployment
The NSA announcement of 1 October 2026 says new commercial National Security Systems must be able to support quantum-resistant algorithms starting in 2027, and that legacy systems unable to do so should be phased out by 2030. The review notes that these are sector-specific dates and say nothing about coverage.
On the quantum side it separates measurements from roadmaps. Google's January result improved logical error by a factor of 2.15 going from distance 3 to distance 5 (the size of the error-correcting code); a separate 15% gain was simulated for a hexagonal layout, and the review warns against conflating the two.
IBM's August news was two cryogenic modules joined and cooled. No source in the batch broke RSA-2048 or deployed elliptic-curve cryptography with a quantum computer.
Every source is tagged by the kind of evidence it is
Each source is tagged as a standard, measurement, first-party claim, policy, forecast, inference or hypothesis. The method note under them states that a source's description of its own findings is not equivalent to independent validation.
Still open
No latency benchmark has been run. The review found no primary evidence that HQC is a final FIPS standard, and the TLS document for ML-DSA is still a draft. The shared-generator weakness is a research question with no result yet.
Sources and notes (10)
- RFC 10024: PQ/T hybrid key agreement for TLS 1.3 — the three hybrid groups and their serialisation
- RFC 9954: Hybrid key exchange in TLS 1.3 — the generic construction and the shared-RNG warning
- NIST SP 800-227: Recommendations for KEMs — final guidance on key-encapsulation mechanisms
- RFC 9881: ML-DSA in X.509 — signatures as a separate migration
- NSA: post-quantum measures for national security systems — the 2027 and 2030 dates
- NIST selects HQC — the March 2025 selection
- Google Research: dynamic surface codes — the logical-error measurement
- IBM: first modular cryogenic systems connected — a roadmap milestone, not cryptanalysis
- Industry review, 14 Oct 01:41 — the notes this was written from
- Industry review, 14 Oct 02:33 — the notes this was written from
Cryptography researchAll 10 parts →
