cryptonist
A bank vault standing open, its door thickness and locking mechanism visible

Wallets

Can a seed phrase be hacked? What the arithmetic says

Not by guessing, if it was generated properly. A 12-word phrase is one of 2128 possibilities, and the arithmetic leaves no room for a search to find one. Phrases are lost in other ways: made from too little randomness, typed into a page that asked for them, read by malware, or photographed into a gallery that something else could see. Each of those has a documented case, and none of them involved breaking the cryptography.

The search an attacker would have to run

A recovery phrase under BIP-39 encodes 128 bits of randomness in 12 words or 256 bits in 24, so there are 2128 valid 12-word phrases, about 3.4 × 1038. Each candidate has to pass the phrase’s checksum, then go through 2,048 rounds of HMAC-SHA512 inside PBKDF2, then through BIP-32 key derivation, and the addresses it produces have to be checked against a chain.

One of the disclosures below gives a measured rate to scale from. Its researchers put a search of 232 candidates, about 4.29 billion, at a few days on an average gaming PC, at most. The 12-word space is 296 times larger. With a billion such machines, and taking “a few days” as three, the full search would take around 6.5 × 1017 years.

An attacker does not need a particular phrase, though. Any funded one would do, and that changes the sum. If there were a billion funded phrases to hit, the expected work before finding one falls to about 298 candidates. That is still 266 times the 232 search, and with the same billion machines it comes to roughly 650 million years. The number of targets would have to be absurdly large before guessing became a plan.

The keys behind the phrase are no softer. BIP-32’s security section assumes that recovering a private key from its public key means solving the elliptic-curve discrete logarithm problem, at about 2128 operations. 12 or 24 words works through what the longer phrase does change; for this question the answer is the same at either length.

So when a phrase is “hacked”, someone has almost always skipped the search. Either the space was never as large as the word count implied, or the phrase was handed to them.

When the randomness was never there

A phrase is only as random as the software that generated it. The words can carry 128 or 256 bits; nothing in them proves the generator supplied that many.

Milk Sad (CVE-2023-39910). Libbitcoin Explorer is an open-source command-line tool whose bx seed command generates wallet entropy. The NVD entry says versions 3.0.0 through 3.6.0 used the mt19937 Mersenne Twister generator in a way that restricts the internal entropy to 32 bits regardless of settings, letting attackers recover any private keys generated from its output, and that this was exploited in the wild in June and July 2023. It also records the vendor’s position that its documentation sufficiently advised against using bx seed, a position others dispute. The researchers’ disclosure found the generator was seeded from 32 bits of system time, so any wallet it produced could be found in about 4.29 billion attempts at most. The two victims they started from had used 24-word phrases created on air-gapped laptops and loaded into hardware wallets. The researchers estimate that over $900,000 of assets were moved in the thefts, at August 2023 exchange rates, while noting that some drained wallets may have been taken through other weaknesses.

Trust Wallet’s browser extension (CVE-2023-31290). The same generator caused an earlier failure. The NVD entry says Trust Wallet Core before 3.1.1, as used in versions 0.0.172 to 0.0.182 of the Trust Wallet browser extension, drew its phrases from 32 bits of entropy, which allowed only about four billion possible mnemonics, and that the flaw was exploited in December 2022 and March 2023.

Coldcard (2026). On 30 July 2026 Coinkite published a security advisory saying a firmware bug had weakened seed generation on its Coldcard devices. The devices, it says, were not hacked or remotely accessed; attackers regenerated the private keys from the weak seeds offline. Seeds from Mk2 and Mk3 firmware 4.0.1 to 4.1.9 are affected, as are seeds generated on the Mk4, Q and Mk5 before the fixed releases, which carry about 72 bits of entropy rather than the expected 128. Updating the firmware does not repair an existing seed. Coinkite does not consider a seed at risk from the bug alone if at least 50 fair, independent, private dice rolls went into it and were never recorded or exposed, since dice input of that size contributed at least 128 bits of entropy by itself. As of 22 November 2026, per DefiLlama, its hacks record lists the incident on 31 July 2026 at $116.0M, classed as a key compromise through weak key generation; it appears among the 2026 incidents.

Profanity (2022). Not a seed-phrase tool, but the same failure. Profanity generated Ethereum vanity addresses, ones with a chosen pattern of characters. In September 2022 contributors to 1inch disclosed that it seeded its 256-bit private keys from a random 32-bit value, and that they had working code to recover the private key of any Profanity address, starting from the public key, which can be recovered from any transaction the address has signed. The tool’s author archived the project the same day, with a note that the affected binaries had been removed.

What these cases share is that the theft happens offline, silently, and can come years after the mistake. Nothing the owner did at the time of the theft could have stopped it, because the exposure was fixed at generation. It is the reason whether anyone outside can check a wallet’s build is a real criterion, and the case how crypto wallets actually fail describes as rare and catastrophic. In DefiLlama’s hacks record, weak key generation is one technique within the key-compromise class, beside compromised private, hot-wallet, deployer and validator keys and keys that were brute-forced.

When the phrase is handed over

The commonest route needs no technical flaw at all. A phrase typed into a page, a form or a chat window has simply been given away.

Canada’s Anti-Fraud Centre, in a bulletin on wallet fraud that New Brunswick’s financial regulator republished in August 2026, lists the channels: fake websites, emails, text messages, phone calls, letters, QR codes and harmful software, often with a claim that something is wrong with the wallet and needs fixing at once. It says that no police service, government agency, financial institution, wallet provider or legitimate support service will ask for the phrase by email, text, social media, online form or letter. MetaMask’s guidance says the same of its own staff, who will never ask for it.

The pages are built to be found. Microsoft’s threat researchers have documented a spoofed MetaMask site whose domain added one word to the real name, phishing sites placed at the top of search results as sponsored adverts, and fake wallet apps that copied the legitimate icon. Someone who has already lost funds is a target a second time: the Canadian bulletin describes recovery fraud, in which a person claiming they can retrieve stolen crypto asks for wallet details or a fee up front. Crypto recovery scams covers that pattern.

When software reads it

Malware has no need to guess a phrase it can read. The same Microsoft research describes wallet-targeting malware, which it calls cryware, and one of its techniques is what Microsoft calls clipping and switching: the malware watches the clipboard for a string shaped like a wallet address and, when the user pastes, substitutes the attacker’s. That redirects a payment rather than stealing a phrase. The report also describes clipboard-monitoring trojans that steal sensitive data a user copies (a phrase copied in order to paste it somewhere is exactly that).

Microsoft also describes cryware that uses text patterns to find seed phrases, its example being twelve words separated by spaces, and cryware that goes after wallet applications’ storage files by name. Keyloggers capture phrases as they are typed, and Microsoft notes that some also take screenshots, which defeats copying and pasting as a workaround.

Substitution can be caught on the confirmation screen, the check set out in how wallets fail. Address poisoning arrives at a similar result by a different route.

When it was photographed

A photograph of a phrase is a copy readable by every app with access to the photo library. In February 2025 Kaspersky reported a campaign it named SparkCat: Android and iOS apps, some distributed through Google Play and Apple’s App Store, carried a hidden component that used optical character recognition to read images in the user’s gallery and send the ones matching keywords supplied by the attackers’ server. The target, Kaspersky says, was recovery phrases for crypto wallets. Kaspersky says the infected Google Play apps had been downloaded more than 242,000 times, that this was the first stealer it had found in the App Store, and that Apple and Google removed the apps after its report. It also cites earlier research by ESET, from March 2023, into modified messaging apps that scanned galleries the same way.

Cloud backup widens that exposure from one phone to one account, held by a company and protected by whatever secures the login. The Canadian bulletin advises against keeping the phrase in an email, cloud account, online document or photograph, and Microsoft’s researchers advise against storing it on the device or in cloud storage at all. Where to store a seed phrase deals with the alternatives.

What the question usually means

“Can my seed phrase be hacked” is, in practice, two questions. The first is whether the phrase was generated with real randomness by software that has been examined. That is settled at creation, and the only fix afterwards is a new phrase. The second is whether it has ever been typed, copied, photographed or entered anywhere other than the wallet that made it, and every route above is a version of that one.

If the answer to either is uncertain, a phrase cannot be rotated like a password. The remedy is a new phrase and a transfer of the funds, as what a recovery phrase controls explains. A BIP-39 passphrase adds a second secret that a leaked or weakly generated phrase does not give away, which is why Coinkite’s advisory treats a strong one as an independent barrier. It does not repair a phrase that was weak to begin with.