Beyond FIDO2: Eliminating the Single Point of Failure in Authentication

FIDO2 security keys and passkeys have made strong, phishing-resistant authentication practical. Together with NTNU, we ran an experiment to learn what it takes to go one step further: to split the signing key across several authenticators, so that no single device holds it and the server sees nothing new. The work has been peer reviewed and accepted at the Norwegian Information Security Conference (NISK 2026), and the report and code are public. Here is what we did, what it cost, and what we learned.

Two challenges, and a different approach

FIDO2 removes passwords, phishing and credential stuffing, and hardware keys and passkeys are the right choice for most users and services. Two challenges remain in any setup where the credential is tied to a single key. The first is availability: when a device is lost, the user needs a recovery path, and that path must be as strong as the authenticator itself. The second is concentration: for the most sensitive accounts, some organisations want no single device, and no single person, to be able to authenticate alone.

Today these challenges are handled well with backup authenticators, synced passkeys and recovery procedures. We wanted to explore a complementary approach that addresses both in the cryptography itself.

Threshold signatures do this. The key is split into n shares on n devices, and any t of them can sign together. Fewer than t shares reveal nothing about the key. Both numbers are policy: t sets how many devices an attacker must compromise, and n minus t sets how many the user can lose. A two-person rule becomes a property of the cryptography instead of a procedure. We used 2 of 3 as the running example, but nothing in the design depends on those numbers.

For the relying party nothing changes. It stores one public key and verifies one ordinary signature, and cannot tell whether one device signed or several.

FROST on authenticator hardware, behind a hostile relay

We implemented FROST, the IETF threshold signature standard (RFC 9591), on the OFFPAD development board and on a more constrained STM32 board, both running Zephyr, the real-time operating system the OFFPAD uses. To our knowledge, this is the first reported implementation of FROST key generation and signing on FIDO authenticator hardware.

The devices cannot reach each other, so a computer relays every message. We treated that computer as an attacker. Secret shares are encrypted end to end, so it cannot read them. Every message is signed by the sending device and checked against a list of device keys set up in advance, so the relay cannot alter messages or insert keys of its own. If a device receives a bad share, the protocol establishes which party cheated. The relay can still block a login. It cannot produce a wrong one.

In concrete terms: an attacker who controls the relay and up to t minus 1 devices, including everything stored on them, can neither recover the key nor sign without an honest device taking part. One caveat: in the prototype, the devices sign whatever the relay sends. Until each signature requires the OFFPAD’s fingerprint check, the design protects the key, not each individual login.

About one second per login

The cryptography is not the bottleneck. A signature share takes about half a second on the device, and combining the shares takes two milliseconds on the host. Flash writes that guarantee one-time values are never reused, together with serialisation and coordination, take up to as much again. Moving a message takes tens of milliseconds over USB or a serial link. In total, one login takes roughly a second. Key setup happens once per credential.

A second is longer than the near-instant tap users know from a single key, so this is not aimed at everyday logins. For administrators of critical systems, duty officers, or anyone whose credential is a high-value target, it is a modest cost for the added protection.

Three gaps to close before production

None of them is the cryptography.

Standards support. FIDO2 identifies signature algorithms by registered identifiers, and the combination of scheme and curve we used has none, so no relying party can accept the credential today. We tested the shortcut of inserting the group key into a standard WebAuthn registration, and it does not work: a threshold authenticator has to implement the authenticator protocol (CTAP) itself. The fix is within reach. Ed25519 belongs to the same family of signatures, it is registered in WebAuthn, and RFC 9591 defines FROST over it. Switching changes the curve, the encodings and the hashing, not the protocol.

Device replacement. The setup survives losing a device, but production also needs a way to enrol a replacement and make the lost share useless, since in 2 of 3 it is half of what an attacker needs. Both require share refresh: new shares for the same key, so relying parties notice nothing. The technique is well known and reuses most of what we built.

Enrolment. Each device must know which other devices belong to its group. In the prototype we configured this by hand. In production it must be established under the user’s or the organisation’s control and then locked on the device. This step decides the security of the whole arrangement, and it is where we would focus next.

Peer reviewed and open

A condensed version, “Implementing Threshold Signatures for FIDO Authentication”, has been accepted for publication at NISK 2026. The full technical report, with the measurements, the threat model, reproduction steps and the complete list of open issues, is available here: Implementing Threshold Signatures for FIDO Authentication. The source code is on GitHub: https://github.com/JavierGarcia22/FROST_Threshold.

This is an experiment, not a product announcement. We ran it to learn what threshold FIDO authentication takes in practice, so that any later decision about production rests on measurements rather than assumptions. We publish the open issues alongside the results because that is how the work should be judged.