Post-Quantum IPsec VPN Migration: ML-KEM in IKEv2, Pre-Shared Keys, and What to Test Before Changing a Tunnel
How to move IPsec VPN tunnels to post-quantum key establishment: where the IKEv2 standards for ML-KEM stand in November 2026, how pre-shared key options compare, why message size matters, and how to catch silent fallback.
Site-to-site and remote access VPN tunnels carry a large share of an agency’s traffic across networks the agency does not control. Most of those tunnels use IPsec, and most set up their keys with the Internet Key Exchange protocol version 2 (IKEv2) using elliptic curve or finite field Diffie-Hellman. Both can be broken by a large quantum computer, so a tunnel recorded today can be decrypted later. We covered that collection risk in our post on harvest now, decrypt later.
Executive Order 14412 sets December 31, 2030 as the date for federal key establishment to move to post-quantum cryptography (see our summary of the federal mandates). For VPNs, the work is more involved than selecting a new algorithm in a menu. This post covers where the standards stand, the three available mechanisms, and what to test before changing a production tunnel.
Where the IKEv2 standards stand in November 2026
The pieces arrived over six years, and some are not final.
- RFC 9370 (May 2023) lets IKEv2 run up to seven additional key exchanges after the first one. The session keys depend on every exchange, so the tunnel stays protected if any one of them holds.
- The IETF document that defines ML-KEM for IKEv2 has been approved and is in the RFC Editor queue. It has no RFC number yet. It registers key exchange identifiers 35, 36, and 37 for ML-KEM-512, ML-KEM-768, and ML-KEM-1024.
- RFC 8784 (June 2020) mixes a pre-shared secret into the key derivation. RFC 9867 (November 2025) revises that approach so the secret is applied earlier in the exchange.
- Post-quantum signatures for IKEv2 authentication are specified in a separate document, also in the RFC Editor queue, covering ML-DSA and SLH-DSA.
For national security systems, authors at NSA’s Center for Cybersecurity Standards maintain a draft IPsec profile for CNSA 2.0. The October 30, 2026 revision calls for a classical exchange first, ML-KEM-1024 as the additional exchange, ML-DSA-87 for authentication, and AES-256 in GCM mode. It is an individual draft without IETF consensus, so treat it as an indication of direction and confirm requirements through NSA’s approval processes.
Product support is still uneven. CISA’s list of product categories (January 23, 2026) places networking hardware, including routers and firewalls, among the categories that are transitioning, not among those where post-quantum products are widely available.
Three mechanisms, and what each one leaves open
| Mechanism | What it protects | What it leaves open | Main cost |
|---|---|---|---|
| Pre-shared key in IKE_AUTH (RFC 8784) | Keys for the IPsec traffic | The initial IKE exchange, including peer identities and traffic selectors, can still be read by a later quantum attacker | Generating, distributing, and rotating a secret for every pair of peers |
| Pre-shared key in IKE_INTERMEDIATE (RFC 9867) | Keys for the IPsec traffic and the initial IKE exchange | Nothing further on confidentiality | The same key distribution work, plus newer software on both peers |
| Additional ML-KEM exchange (RFC 9370) | Keys for the IPsec traffic and everything after the first exchange | Authentication stays classical until post-quantum signatures are deployed | Larger messages, fragmentation support, and updated software on both peers |
Two details from RFC 8784 belong in any configuration standard. The secret should hold at least 256 bits of entropy, and one derived from a password or a name is not secure. The RFC also recommends deploying the secret as optional while peers are upgraded, then making it mandatory, because an optional setting lets the tunnel proceed without it.
Pre-shared keys are useful where a device cannot run ML-KEM yet. The ML-KEM exchange is the long-term path because it removes the need to move secrets between sites.
Message size is the main engineering constraint
The ML-KEM document for IKEv2 lists the payload sizes that each side sends.
| Parameter set | Initiator payload | Responder payload |
|---|---|---|
| ML-KEM-512 | 808 bytes | 776 bytes |
| ML-KEM-768 | 1,192 bytes | 1,096 bytes |
| ML-KEM-1024 | 1,576 bytes | 1,576 bytes |
The first IKEv2 exchange, IKE_SA_INIT, has no fragmentation mechanism of its own. The document therefore says implementations using UDP should not place ML-KEM-768 or ML-KEM-1024 in that first exchange where the path’s maximum packet size is unknown or restricted. The expected pattern is a small classical exchange first, followed by ML-KEM in an encrypted IKE_INTERMEDIATE exchange, where IKEv2 fragmentation (RFC 7383) can split the message.
An ML-KEM-1024 payload is larger than a standard 1,500-byte Ethernet frame before any headers are added. For a tunnel that must use that parameter set, fragmentation support on both peers is a requirement. The same applies later to authentication: an ML-DSA-87 signature is 4,627 bytes under FIPS 204, and the signature document requires fragmentation support for that reason.
A tunnel that comes up may still be classical
During a rollout, gateways usually offer both a proposal with ML-KEM and one without, so that peers not yet upgraded keep working. If the peer selects the classical proposal, the tunnel establishes normally and the status page shows it as up.
The ML-KEM document addresses this. A responder that knows its initiators support ML-KEM should reject proposals that lack it, and an initiator in the same position should offer only proposals that include it. That is a configuration decision per peer, and someone has to make it after each peer is upgraded.
To confirm a tunnel has migrated, read the key exchange methods negotiated for each security association from the gateway. Tunnel state alone does not show it.
An illustrative example: 140 tunnels
Consider a hypothetical agency with two hub gateways and 140 branch tunnels. All numbers are illustrative.
An inventory of tunnel configurations finds four groups that need work before ML-KEM can be enabled:
- 9 tunnels still use IKEv1, which the IETF deprecated in RFC 9395. None of the mechanisms above exist for IKEv1, so these move to IKEv2 first.
- 31 branch routers run a software release without the ML-KEM identifiers. A later release adds them, and these routers need a memory upgrade to run it.
- 12 tunnels cross satellite links with a path limit near 1,280 bytes.
- 6 tunnels terminate at a partner organization whose gateway uses an early, vendor-specific identifier for its post-quantum exchange.
The remaining 82 tunnels can be changed by configuration alone.
The agency sequences the work. The 82 ready tunnels move first, in batches, with the hub set to require ML-KEM for each peer once its batch is verified. The 31 older routers receive 256-bit random pre-shared keys under RFC 8784 as an interim measure, recorded with an end date tied to the hardware upgrade. The satellite tunnels are tested in a lab at the same path limit, which shows one router model that does not negotiate IKEv2 fragmentation by default. The partner tunnels wait for both sides to agree on the registered identifiers.
Twelve weeks in, a report drawn from negotiated parameters shows 79 tunnels using ML-KEM. Three in the first batch are up but classical, because their hub entries were never changed to require it.
Questions for suppliers and change reviews
- Which RFC 9370 key exchange identifiers does the release support: the registered values 35 to 37, or vendor-specific ones?
- Is ML-KEM-1024 among the available parameter sets?
- Is IKEv2 fragmentation enabled by default, and has it been tested at the smallest path size in our network?
- Does rekeying repeat the ML-KEM exchange, or only the first negotiation?
- Is the ML-KEM implementation inside a validated cryptographic module, and is it used when the device runs in approved mode?
- Can the gateway report the negotiated key exchange methods per security association, in a form we can export?
Keeping evidence per tunnel
VPN configurations change often. Peers are replaced, templates are reapplied, and a proposal list that was tightened can be loosened during an outage. The useful record is which algorithms each tunnel negotiated, kept current.
Sentinel helps organizations understand and change cryptography across many systems at once, which includes finding where key exchange settings live across gateways and templates. CAM provides asset-level cryptographic evidence for assets in production after migration, so a statement such as “these tunnels negotiate ML-KEM” can be checked per asset. Both products have a provisional patent issued.
If your program is planning a VPN migration or writing post-quantum requirements for network equipment, contact QuantumWorks.