Mike Belshe says real AI hacking claims should be tested against institutional Bitcoin custody, not flawed sandboxes. #
BitGo's chief executive has challenged artificial intelligence developer Anthropic to steal about $6 million in bitcoin from one of the company's wallets, after Anthropic disclosed that its Claude models gained unauthorised access to real organisations during cybersecurity testing.
Mike Belshe, a software engineer who co-founded BitGo and helped develop the HTTP/2 web protocol, transferred exactly 100 bitcoin into a publicly visible wallet on Friday before posting its address on X alongside a challenge directed at Anthropic.
The challenge followed Anthropic's publication of an internal review describing three incidents in which Claude models accessed live systems after a testing environment was mistakenly left connected to the public internet.
Belshe argued those incidents reflected failures in sandbox configuration rather than evidence that current AI systems could compromise institutional-grade cryptocurrency custody.
Belshe Publishes Wallet Address #
Belshe posted the wallet address on X with a direct challenge to Anthropic. 'Either @AnthropicAI is terrible at building sandboxes... or excellent at marketing. (or both),' he wrote. 'But enough with the "hacking monster" games. Do it for real. I put this in a @BitGo wallet for you. Go get it.'
The wallet contained exactly 100 bitcoin, valued at roughly $6 million at prevailing market prices. Blockchain records showed the funds remained in the address following the challenge. No outgoing transactions had been recorded as of Sunday.
Anthropic has not publicly responded.
Anthropic Report Described Real-World Access #
Anthropic published its findings on Thursday after reviewing more than 141,000 cybersecurity evaluation runs involving Claude models.
The company said three incidents, spanning six evaluation sessions, resulted in models reaching real organisations after external testing partner Irregular mistakenly left evaluation environments connected to the internet.
According to Anthropic, the models had been instructed they were operating inside isolated capture-the-flag exercises and believed the external systems formed part of those evaluations.
In one incident, Claude Opus 4.7 located a real website sharing the name of a fictional target, identified exposed services, recovered credentials and accessed a production database.
Anthropic said the models did not intentionally escape their environments. Instead, they followed instructions inside evaluation systems that had been improperly configured. The company said it notified the affected organisations and introduced additional safeguards after discovering the incidents.
Institutional Custody Draws Attention #
Belshe's criticism centred on the difference between exploiting poorly isolated testing environments and accessing institutional cryptocurrency custody infrastructure.
BitGo provides custody services using multi-signature and multi-party computation technologies designed to distribute signing authority across multiple independent cryptographic keys rather than relying on a single credential.
Belshe did not claim the platform was impossible to compromise. Instead, he argued that successfully moving funds from a production-grade custody system would provide a more meaningful demonstration of AI hacking capability than incidents caused by testing errors.
Wallet Remains Untouched #
The challenge has attracted attention across the artificial intelligence and cryptocurrency industries because it places a publicly verifiable financial target at the centre of an ongoing debate over AI cybersecurity capabilities.
While Anthropic's report documented real-world access during evaluation exercises, no publicly available evidence indicates that any AI system has independently compromised an institutional cryptocurrency custody platform. As of publication, the bitcoin remained in the BitGo wallet and Anthropic had not commented on Belshe's challenge.
© Copyright IBTimes 2025. All rights reserved.