diff --git a/proposals/3834-tofu.md b/proposals/3834-tofu.md
new file mode 100644
index 000000000..08c3031bf
--- /dev/null
+++ b/proposals/3834-tofu.md
@@ -0,0 +1,144 @@
+# MSC3834: Opportunistic user key pinning (TOFU)
+
+In the Matrix spec at present, there are no real provisions for
+properly authenticating devices belonging to other users, aside from
+the fairly rare cases where users actually take the time to verify
+each other out-of-band. Most clients encourage users to cross-sign
+their own devices, but this has little practical value for most users
+in the ecosystem at present, unless you're willing to verify all of
+your contacts or trust the user keys reported by homeservers.
+
+The goal of this proposal is to establish a TOFU-style mechanism where
+users' signing keys are pinned when they're first encountered;
+internally, this will look similar to user verification, but will
+require no action on either user's part, and correspondingly will not
+be considered as high a level of trust. Nevertheless, this at least
+raises the bar for Mallory-in-the-middle attacks by malicious
+homeservers, as they can no longer quietly falsify signing keys for
+users who are already in contact with each other.
+
+## Proposal
+
+This proposal adds a fourth ed25519 keypair to a user's set of
+cross-signing keys: the TOFU Signing Key (TSK), identified by
+`m.cross_signing.tofu_signing`. This key will be uploaded and stored
+alongside the MSK, SSK, and USK; it will be published to the
+`keys/device_signing/upload` using the new optional field
+`tofu_signing_key`.
+
+Whenever a user Alice and a user Bob are in a room together for the
+first time, Alice's device will automatically and silently cross-sign
+Bob's MSK using her TSK. This effectively pins Bob's MSK; in future
+interactions, Alice will be able to check that any device of Bob's is
+signed by Bob's SSK which is signed by the pinned MSK, confirming that
+the device really does belong to Bob as long as that very first
+interaction wasn't tampered with.  If Alice's device later observes an
+MSK for Bob that differs from the one signed by her TSK, the device
+may warn Alice that Bob's key has changed, and could potentially
+refuse to share keys with devices that aren't signed under the
+original SSK/MSK until Alice chooses to accept the new key.
+
+After a device signs one or several new MSKs with the TSK (on joining
+a new room, or observing a new member join a room), it should upload
+these new signatures to the homeserver via the
+`keys/signatures/upload` API endpoint. However, it is important to
+ensure that the homeserver cannot perform a denial-of-service attack
+on key pinning by simply dropping TOFU signatures.  For this reason, a
+user's devices should periodically ask each other to confirm a hash of
+the complete set of all cross-signing signatures created by that
+user. At any point, a device may send any or all of the user's other
+devices an encrypted to-device message, with event type
+`m.signatures_hash_request` and no fields. When the device receives
+the message, it should reply with an encrypted to-device message with
+event type `m.signatures_hash` and one field `sha256`. This field
+contains a SHA256 hash of the canonical JSON encoding of the complete
+set of cross-signing signatures known to that device, in the format
+for uploading to `keys/signatures/upload`, excluding any unsigned
+fields - that is, a map from user ID, to key ID, to a signed JSON
+object for that key.
+
+## Potential issues
+
+If a client does choose to warn users every time a pinned MSK changes,
+users in many large rooms could be overwhelmed with warnings from
+random people losing their devices and performing key resets.  There
+are several things clients can do to alleviate this. First, rather
+than popping up a warning immediately in a toast or dialog as soon as
+a key changes, clients should only warn the user when they actually
+open one of the rooms where the changed key is present. In addition,
+users should be given the option to automatically accept all future
+key changes in a particular room, and the option to accept all key
+changes globally if that's in line with the user's threat
+model. Clients could still display a small state-event-like message in
+room timelines notifying the user of a key change, or could simply
+accept all key changes silently.
+
+## Security considerations
+
+In the malicious homeserver scenario that this proposal aims to
+mitigate, the homeserver could attempt to simply exclude TSK
+signatures from its responses to key queries, preventing the
+propagation of pinned keys between devices, and possibly even causing
+a device to lose track of pinned keys over time.  To prevent this,
+it's important that clients maintain a persistent record of TSK
+signatures without relying on the homeserver to store that
+information. Signatures must never be removed from this record, even
+if they aren't present in homeserver responses. However, if multiple
+valid signatures for the same key from the same TSK have been created
+due to race conditions, only the signature whose base64 representation
+comes first lexicographically should be kept.
+
+When devices ask each other to confirm the hash of the set of
+cross-signing signatures, they may not receive a response; this could
+be because the other device is offline, it could be because the device
+is using an older version of the spec and doesn't understand the
+request, or it could be because the homeserver is performing a
+denial-of-service attack. If a device doesn't get a response, and does
+see other activity from the other device, it could possibly eventually
+display a warning encouraging the user to upgrade their other
+device. If the device observes too many missing key indices in its
+to-device Olm session, however, that may be a sign that the homeserver
+is dropping messages.
+
+In this threat model, another thing the homeserver can do is falsely
+claim that the user's cross-signing keys have been reset by another
+device. For the most part, this simply amounts to a denial-of-service
+attack on cross-signing, which is difficult to prevent in
+general. However, it's important to note that if a formerly-verified
+device is told by the server that there's been a key reset, it must
+still hold onto its record of signatures created with the old TSK, and
+re-sign those with the new TSK once it has been able to confirm the
+new MSK as part of becoming re-verified.
+
+## Unstable prefix
+
+This proposal does not add any new API endpoints, but does add some
+new fields and message types to existing data structures. These field
+names and message types will be prefixed with `org.matrix.msc3834.v1`.
+Test implementations can identify the TSK as
+`org.matrix.msc3834.v1.cross_signing.tofu_signing`, use the optional
+field `org.matrix.msc3834.v1.tofu_signing_key` for the
+`keys/device_signing/upload` endpoint, and use the message types
+`org.matrix.msc3834.v1.signatures_hash_request` and
+`org.matrix.msc3834.v1.signatures_hash` for exchanging hashes of the
+set of cross-signing signatures.
+
+## Migration notes
+
+When a device with this functionality first comes online, if the user
+doesn't already have a TSK and if the device possesses the private
+MSK, it should create a new TSK, sign it with the MSK, and publish it
+to `keys/device_signing/upload`. This endpoint requires
+authentication, and therefore the user may need to be prompted for a
+password when they open their upgraded client. If the user already has
+server-side secret storage set up, the device should also store the
+private TSK encrypted in secret storage.
+
+If the user does already have a TSK, devices should obtain the private
+key either from secret storage or by requesting the secret from other
+devices. In the event of a race condition where multiple devices have
+upgraded and created TSKs, the TSK whose base64 representation comes
+first lexicographically will take precedence. Once other devices have
+received the private key for the winning TSK, they should use it to
+re-sign and re-submit any TOFU signatures they've already created, and
+replace any signatures by the old TSK in their persistent records.