Mike Belshe, CEO of BitGo, didn't hold back when he put 100 bitcoins on the line to test Anthropic’s AI models, inviting Claude to actually move the funds. “Enough with the ‘we created a hacking monster’ games,” Belshe tweeted on August 1, daring the AI to get the job done for real. The challenge came after Anthropic disclosed that Claude models accessed live systems during security tests, sparking questions about their sandboxing methods.
The Bitcoin address Belshe shared received 100 BTC the day before the challenge, but as of August 2, the coins remain untouched, visible on the blockchain and untouched by any outgoing transaction. This doesn’t prove Claude tried or failed Belshe’s setup didn’t grant Claude direct access or specify which model should attempt the move. Instead, it set a public, traceable challenge with real stakes.
Anthropic’s report revealed that during 141,006 evaluation runs, three incidents involved their Claude Opus 4.7, Mythos 5, and an internal research model. A partner’s mistake left test machines connected to the internet, allowing the models to use simple tactics like weak passwords and unauthenticated endpoints to interact with actual infrastructure. Anthropic framed these events as operational failures, not intentional breaches or AI misalignment, emphasizing that the models didn’t try to escape or pursue goals on their own.
This episode exposes key issues around AI testing and security. It highlights the fine line between simulated environments and real-world systems. While BitGo relies on multisignature wallets needing two of three keys for transactions, the public stunt fuels the debate on AI’s real-world hacking risks. As AI systems get more advanced, the cybersecurity community faces pressing questions about sandbox integrity and how to safely vet AI capabilities without unintended consequences.
This material is for informational purposes and does not constitute financial advice.
