Can AI Hack Bitcoin? BitGo CEO Dares Anthropic’s Claude to Crack 100 BTC Wallet

BitGo CEO Mike Belshe has challenged Anthropic’s Claude artificial intelligence model to seize 100 Bitcoin from a publicly disclosed wallet, turning growing fears about autonomous AI hacking into a multimillion-dollar experiment.

Belshe published the Bitcoin address on X and invited Claude to move the funds. The wallet still showed no outgoing transactions as of Aug. 2, leaving approximately 100 BTC, worth more than $6 million at current prices, under BitGo’s control.

The challenge followed Anthropic’s disclosure that Claude models had reached the public internet during cybersecurity evaluations and gained unauthorized access to systems belonging to three real organizations.

Belshe dismissed the incidents as evidence of poor sandbox design or clever marketing rather than proof that Anthropic had created a powerful “hacking monster.”

New Trending Crypto Wallet Offers

Sponsored

Disclosure

We sometimes use affiliate links in our content, when clicking on those we might receive a commission at no extra cost to you. By using this website you agree to our terms and conditions and privacy policy.

#1

Safepal

Opened in 2018

Promotions

Trusted, Secure & Crypto Friendly

Coins

+87

Claim Offer

#2

Tangem

Opened in 2017

Promotions

Receive Up to $10 in BTC when you buy and activate a Tangem Wallet.

Coins

+68

Claim Offer

#3

BitBox

Promotions

Trusted, Secure & Crypto Friendly

Coins -

Claim Offer

Anthropic Says Claude Compromised Real Organizations

Anthropic uncovered the incidents after reviewing 141,006 cybersecurity evaluation runs conducted with Irregular, one of its external testing partners.

The company found six problematic runs across three incidents. A configuration error gave Claude internet access even though its instructions described the environment as an isolated simulation. The models consequently treated real online systems as components of capture-the-flag exercises.

Claude used relatively basic methods rather than discovering sophisticated vulnerabilities. These included weak passwords, exposed credentials, unauthenticated endpoints, and SQL injection.

One model accessed a production database containing several hundred records. Another created a malicious Python package and uploaded it to the public PyPI repository.

Fifteen systems downloaded the package before PyPI’s security controls removed it.

Anthropic said one older model continued attacking after recognizing signs that it had reached a real production environment. Its newest research model stopped after concluding that the target existed outside the simulation.

The company attributed the incidents primarily to operational and testing failures rather than a model deliberately escaping containment or pursuing an independent objective.

Why Stealing Bitcoin Presents a Different Challenge

Claude’s ability to exploit misconfigured servers does not mean it can break Bitcoin’s underlying cryptography.

To move Belshe’s 100 BTC, an attacker would need the private keys controlling the wallet. A public Bitcoin address reveals where the funds sit but does not expose the secret information required to authorize a transaction.

Modern Bitcoin security relies on cryptographic problems that conventional computers cannot solve through brute force within any practical timeframe.

AI models may improve vulnerability research, automate phishing or identify mistakes in wallet software, but they cannot simply calculate a properly generated private key from its public address.

The most realistic attack would therefore target BitGo’s security implementation rather than Bitcoin itself.

Claude could theoretically search for software vulnerabilities, leaked credentials, weak access controls, or human errors surrounding the wallet.

BitGo specializes in institutional custody and uses technologies such as multisignature authorization and distributed key management. Those controls aim to ensure that compromising one system does not provide enough information to move client funds.

Challenge Tests Custody More Than Bitcoin

Belshe’s challenge makes a dramatic point, but it does not create a controlled scientific test of Claude’s capabilities.

Anthropic would need to provide the model with internet access, tools, substantial computing resources, and explicit authorization to attack BitGo.

The company’s public safeguards would almost certainly block a request to steal cryptocurrency.

https://twitter .com/AnthropicAI/status/2082965101083320543

The absence of an outgoing transaction, therefore, does not prove that AI agents pose no cybersecurity threat. It instead demonstrates the difference between attacking vulnerable infrastructure and defeating a cryptographically secured Bitcoin wallet.

AI could already help criminals scale phishing, malware development, and credential theft.

However, Belshe’s untouched 100 BTC suggests that even advanced models still need a weakness to exploit. Without a leaked key, implementation flaw, or human mistake, Claude faces mathematics, not merely another poorly protected server.