Zenpanik
The key that stays and the key that travels: where a passkey lives on an iPhone and a MacBy Pavel Nikolov, November 12, 2026

Part 4: One sign-in, two keys

Part 4 of 5 in a deep dive on where a passkey lives on an iPhone and a Mac, and what the Secure Enclave, Touch ID and Face ID do for it. Start with Part 1 if you are new to the series.

One sign-in, step by step

Take the commonest case. You open your bank’s site in Safari on an iPhone, tap the sign-in button, glance at the screen, and you are in. Six things happened, and the two keys took turns.

Website1. ChallengeThe site sends a random number and asks for the passkey of this account.
Keychain2. LookupThe system finds the passkey row for this domain, using the cached metadata key.
Enclave3. Face IDYou glance at the phone. The enclave compares the scan with the enrolled face.
Enclave4. UnwrapThe enclave checks the row’s access rules and releases the key for its secret value.
Keychain5. SignThe passkey’s private key signs the challenge together with the site’s domain.
Website6. VerifyThe site checks the signature with the public key it saved at registration.
The key that staysThe key that travelsThe website
Figure 13. A passkey sign-in on an iPhone with Face ID. The enclave acts in steps 3 and 4 and the passkey acts in step 5. Schematic, assembled by me from About the security of passkeys, Keychain data protection and Uses for Optic ID, Face ID, and Touch ID. Apple does not document which processor performs step 5.

Step 5 is ordinary public-key signing. If sk is the passkey’s private key and pk the public key the site holds, the device computes, in simplified form, s = Sign(sk, challenge, domain) and the site accepts if Verify(pk, challenge, domain, s) is true. The challenge is fresh each time, so a recorded signature cannot be replayed. The site stores only pk, which Apple notes is not a secret, so a breach of the site’s database gives an attacker nothing to sign with.

The domain in that formula is what defeats phishing. NIST describes a passkey as a key pair whose use is “constrained to the domain in which it was created”. A fake page at a look-alike address is a different domain. The system finds no passkey row for it in step 2, so there is nothing for you to approve. A password has no such check. You can type it into any page that asks.

The staying key never signed anything the website saw. Its work was steps 3 and 4: decide that the right person is present, then open the box. The traveling key did the signing. Face ID supplied one bit, yes or no, and supplied it to the enclave only. This is Way A from Part 2. With an enclave-made key, step 5 would move inside the chip.

What each attacker needs

The same six steps give a way to compare attackers. For each one I list what they already hold and what they still need to produce a valid signature in step 5.

AttackerHoldsStill needsOutcome
Phishing siteA convincing copy of the sign-in pageA passkey registered for its own domainFails. No passkey is offered for the wrong domain
Breached websiteThe database of public keysThe private keyFails. Public keys cannot sign
Thief, locked iPhoneThe device and its encrypted keychain fileThe passcode within 10 tries, or a false biometric match within 5About 1 in 100,000 against a random six-digit passcode (my arithmetic, Part 2)
Thief who saw the passcodeThe device and the passcodeNothing more, unless Stolen Device Protection is onCan use saved passkeys. With the protection on and the phone away from familiar places, needs your face or finger
Apple Account takeoverThe account password and the SMS codeA device passcode within 10 tries at the escrow HSMsBy Apple’s description, the keychain stays closed without the passcode
Boot ROM exploit on an older chipThe device, and code running before iOS loadsA way through the enclave and its attempt counterThe usbliter8 authors say user data is not directly exposed
Malware in control of an unlocked deviceThe running operating systemYou, approving a Face ID prompt you believe is realCould misuse an unlocked passkey. My inference from the design. I know of no confirmed public case, and Part 2 lists one unconfirmed claim

Sources for the table: NIST SP 800-63B-4, Appendix B, About the security of passkeys, The Secure Enclave, About Stolen Device Protection, Escrow security for iCloud Keychain, SecurityWeek on usbliter8.

The first two rows are where passkeys beat passwords outright, and the enclave plays no part in either. The protection comes from the protocol. The enclave matters from the third row down, where the attacker has your hardware or your account.

The passcode is the hinge

Rows three, four and five end at the same place. The locked phone falls to the passcode. The account takeover stalls for lack of the passcode. The thief who has the passcode needs nothing else. Part 2 explained why: Touch ID and Face ID are shortcuts to the passcode, and by default the passcode works in their place.

Apple addressed the fourth row directly. Stolen Device Protection, which requires iOS 17.3 or later, is built for someone who has stolen your iPhone and knows your passcode. When the phone is away from familiar locations such as home and work, some actions need Face ID or Touch ID with no passcode alternative. Using passwords or passkeys saved in the keychain is on that list. Changing the Apple Account password or the passcode itself needs a biometric match, an hour’s wait and a second match.

With this setting the biometric stops being a shortcut and becomes a requirement. With it on, a watched passcode no longer opens the passkeys. By default the extra checks apply only away from familiar locations, and Apple offers an “Always” option that applies them everywhere.

Synced or device-bound

A device-bound passkey is one whose private key cannot leave a single piece of hardware. A USB or NFC security key holds one. So does any key created inside the enclave under the rules in Part 2. Chrome’s own Touch ID credentials on a Mac are of this kind in the Chromium revisions I read. Apple’s own Passwords flow lets a user save a passkey to a hardware security key instead of the keychain. The trade is direct. A device-bound key has no copies to attack and no copy to fall back on. NIST grades sign-in methods on three authentication assurance levels, AAL1 to AAL3, and the table uses them.

Synced passkey (iCloud Keychain)Device-bound passkey (security key)
Resists phishingYesYes
Copies of the private keyOne per device in the circle of trust, plus the escrow recordOne
If the device is lostStill on your other devices, or recoverable from escrowGone. You need a second registered key or the site’s recovery process
Ways in for an attackerAny device passcode in the circle, or the account plus a passcode within 10 triesThe physical key and its PIN or touch
NIST assurance levelUp to AAL2Can meet AAL3, which requires a non-exportable key
Proof of hardware to the websiteNone for consumers, according to Corbado: Apple sends no attestation for iCloud Keychain passkeysAvailable through attestation on many keys

Sources for the table: Secure keychain syncing, Escrow security for iCloud Keychain, NIST Appendix B, NIST assurance levels, Corbado on iOS 26. The security-key column is my general description and varies by product.

Attestation is a signed statement from the hardware about what kind of device created a key. A bank that wants proof a passkey sits in a particular chip would ask for it. Corbado reports that for consumer iCloud Keychain passkeys Apple sends a zeroed-out identifier and keeps enterprise attestation for managed devices. NIST accepts that. Its guidance says the lack of attestation “SHOULD NOT block the use of syncable authenticators for broad public-facing applications”. So, on Corbado’s account, a website receiving a consumer iCloud Keychain passkey learns that the signature is valid and learns nothing about where the key is kept.

A Taleb lens: pebbles, a stone, and a removal

Nassim Taleb’s test for fragility fits this design closely. In his 2013 letter to Stewart Brand he states the principle that anything fragile is harmed more by one stressor of intensity X than by k stressors of intensity X/k. If H is the harm, H(X) > k × H(X/k). His picture is a large stone against the same weight in pebbles.

Passwords that are unique and kept in your head fail as pebbles. One phished password costs one account, and the next phish costs another. A keychain fails as a stone. One learned passcode on one device, or one completed recovery, exposes every passkey and password in it together. Any password manager has this shape, so it comes from keeping credentials together, and sync adds more devices that can open the store.

Unique passwords: losses arrive one at a timeThree phished passwords, three accountsOne keychain: one event, every account72 credentialsOne passcode learned, or one recovery completed
Figure 14. The same 72 credentials under two failure patterns. Illustrative: the counts are mine and stand for no measured data. The principle is from Taleb’s letter to Stewart Brand.

I think the harm is worse than proportional. Losing 72 accounts in one event includes the email account that resets the others and the bank account that funds the recovery, so the victim loses the means of repair along with the accounts. Taleb’s first precautionary rule in the letter is to split sources of harm as much as possible, and his third is to keep variations small and independent so that errors do not spread. A second, device-bound key on the few accounts that matter most is that rule applied to sign-in.

The letter also argues that we cannot predict the effects of adding something new, while we can predict the effect of removing something that was not historically part of a system. His examples are substances such as tobacco, so I am stretching the idea. Passkeys do one of each. They remove the secret that the server had to store and the user could type into the wrong page, and the effect is the predictable one in the attacker table’s first two rows. They add sync, escrow and export, and those are new machinery with a short record.

On records, Taleb writes that time has a lot of value in detecting what is breakable and what is not. The enclave has had thirteen years of attack, and passkey export has had one. A clean record is also the evidence of Taleb’s turkey, whose confidence in the farmer rises with every day it is fed. Apple’s design does not lean on the record. The counters in the chip and in the escrow HSMs fail closed, by erasing. The longer-term stone is mathematical. Passkeys commonly use elliptic-curve signatures, which a large quantum computer would break.

For most accounts the synced passkey is the right trade, because losing access is the likelier harm. For the few where a single event would be unrecoverable, I would add a key that stays.

Continue to Part 5: What to do