# Keplr and Web3Auth: $6.01 million in dispute over security

> Bitquery counted $6.01 million leaving 44,041 Keplr wallets created with Google. Web3Auth says they were rescues. The figure is still unresolved.
> By José Maldonado · Cybersecurity · 2026-11-09

Bitquery counted $6.01 million leaving Keplr wallets during the night of November 6-7, 2026. The affected wallets had something in common: they had been created with the Google or Apple button, not with a twelve-word phrase. Web3Auth, the company that provides the lock behind those buttons, maintains that much of that money was not stolen by anyone, but rather moved by the company itself to keep it safe.

The distinction decides whether we are talking about a six-million-dollar theft or a scare with almost everything recovered. A wallet with social login works like this: the user signs in with their Gmail account and the provider stores part of the private key, the code that authorizes transactions. It is convenient and eliminates the secret phrase that almost nobody writes down. It also concentrates risk, because if the shared piece fails, thousands of wallets fail at once.

Web3Auth has been part of MetaMask since Consensys integrated it, so the flaw touched the infrastructure of one of the most widely used wallets in the sector. Both companies asked users to ignore recovery offers outside their channels, and Keplr recommended not depositing new funds. As of November 9, 2026, no one has published a definitive per-user tally of losses, and that gap is the heart of the story.

## A flaw that puts the platform in jeopardy

Web3Auth confirmed the flaw on November 6 and paused its embedded wallet service that same day, at around 21:13 UTC, a time recorded in the technical account of the pause. The following day, Keplr disabled the Google and Apple buttons to stop the bleeding.

The sequence is short: the provider shuts down, the wallet closes access, users discover empty balances. Hours passed between the notice and the cutoff, and in those hours almost all the money now in dispute was moved.

Josh Lee, co-founder of Keplr, attributed the breach to a Web3Auth flaw and not to his own wallet's software. His estimate puts the scope at close to 20% of wallets created with social login, though he specified that only a fraction had been exploited. Wallets with a recovery phrase or a hardware device were outside the incident, a detail that separates two ways of understanding custody.

Keplr recommended that affected users not deposit new funds and move whatever remained to safe accounts. This is no minor advice, because if the shared key remains exposed, any subsequent deposit falls into the same hole. The [constellation of actors that underpins a wallet's security](https://observatorioblockchain.com/en/sociological-analysis/la-constelacion-de-la-seguridad-dinero-en-tiempos-de-hackeos/) explains why an external link can bring down the whole even when the user does everything right.

## 44,041 wallets, the map of the damage

Bitquery's tracing, by a blockchain data firm, counted 44,041 affected addresses in the Cosmos ecosystem and 7,603 in Injective. Nearly 86% of the wallets touched across seven connected networks recorded transfers after Keplr's first notice. The [investigation of the flows of November 6 and 7](https://bitquery.io/investigations/keplr-wallet-hack) adds the breakdown by network. Cosmos Hub accounts for about $3.50 million. Osmosis adds $1.04 million and Injective adds $615,000.

Two figures in that tally say more than the total. The median loss per wallet was $1.34 and the largest single transfer was around $325,000. Most lost pocket change; a few lost a house.

The detail that most unsettles those who stake is another one. Bitquery documented about 1.96 million ATOM in staking that was moved without completing the 21-day unbonding period, through liquid staking mechanisms that turn the position into a transferable token. With that trick, about $251,000 in rewards changed hands. The promise that staking imposes a three-week wait was broken through the back door.

The list of affected networks also tells a story about concentration. Keplr is not the only wallet in the ecosystem, but it is the most widely used gateway in Cosmos, and a single integrated provider gave one flaw the reach of eleven networks. Injective shows it clearly: a different team suffered the same flaw because it shared a provider.

## The $6.01 million that the company considers false

The figure that circulated first came from adding up everything that left the wallets during the incident window, at November 7 prices. Web3Auth replied that this sum lumps two different things together: thefts and rescues. In its version, part of the transfers were defensive recoveries carried out by the company and its partners to addresses they control, and the [Web3Auth response](https://x.com/Web3Auth/status/2119114303907196932) with the list of verified addresses arrived on November 8 with eleven Cosmos networks and eight EVM chains.

The company quantifies its operation at $5.6 million moved to twelve recovery addresses. If the money is in those addresses, its argument is that a thief does not have it. Bitquery's figure counts outflows, not destinations.

No one outside the company can confirm what was in each wallet before the incident, and the user who sees their balance at zero cannot distinguish a rescue from a theft. Keplr identified recovery addresses that matched some of the receiving wallets examined by Bitquery, so both accept that there were rescues and argue over how many and for how much.

The precedent of the Binance haul serves as a mirror on a smaller scale. The [$3.11 billion that remain still in addresses from the attack](https://observatorioblockchain.com/en/crypto/elliptic-botin-binance-3110-millones-sin-mover/) made it clear that the figures for an incident are measured at several different moments and that none matches the next. Here the same happens with six million and a 48-hour window.

## Why does a rescue look like a theft on the blockchain?

A defensive recovery needs the same key as an attack. If the problem is that the private key was exposed in a third-party service, the way to get the funds out before the intruder is to use that exposure first. On the blockchain, the resulting transaction is identical: the balance leaves the user's wallet for an address they do not control.

Hence the confusion, and also the opportunity. Anyone who empties a wallet can call what they did a rescue, and only the publication of the addresses allows one thing to be separated from the other.

Blockchain analysis sees movements, not intentions. That limitation has two uncomfortable consequences. The first is that hot estimates tend to exaggerate the damage, as the company maintains. The second is that an attacker can hide behind the rescue narrative and gain days while affected users wait for instructions. The only available proof is a list of addresses published by one of the parties.

The fight over the narrative is not new. The [clash between the OKX founder and CZ over downplaying a theft](https://observatorioblockchain.com/en/cybersecurity/el-fundador-de-okx-acusa-cz-minimizar-robo-binance/) showed how quickly an incident turns into a debate over who tells it worse. Something similar is happening with Keplr and Web3Auth, with one useful difference: here both parties publish data that the user can verify.

## What separates a social wallet from self-custody

The practical difference lies in who can move the money without the user's permission. With a seed phrase or a hardware device, the key lives with the owner and the provider has nothing to sign. With social login, part of the key lives on someone else's server and the user never sees it. Convenience and risk are the same piece.

That does not turn self-custody into a refuge. A lost phrase or a badly approved signature empties a wallet just as quickly, with no company to claim against and no list of addresses to consult.

What changes is where the failure lands. In self-custody the error is the user's and the scope is one wallet; with delegated custody the error is the provider's and the scope is tens of thousands. The incident does not prove that one option is better, but rather that risk moves location when the key is outsourced, and that the user rarely knows how much of the key they have handed over.

For these days, the advice from both companies is prudent: do not deposit funds into the affected wallets, move whatever remains and distrust any recovery offer that does not come from their channels. The open question is the usual one: how much of the 6.01 million will return to its owner, and by when.

The answer will not come from a message on social media. It will come from reconciling two lists that already exist, the one of what left the wallets and the one of what rests in the verified addresses, and from accepting that until that subtraction is published, the label of theft and the label of rescue describe the same movement.
