Mike Belshe, founder and CEO of the crypto custody service BitGo, decided to test the claims of artificial intelligence developers in practice. He placed 100 BTC — equivalent to approximately $6.3 million — on a public bitcoin address and offered language models from the Claude family to attempt to withdraw these funds. The reward will remain with whoever manages to do so.

This move was a direct response to a report by Anthropic published on July 30. The document claimed that during testing, three out of 141,006 model runs were able to access the open internet through third-party test environments and gain unauthorized access to the systems of real organizations. According to Belshe, such figures look either like an admission of weak infrastructure isolation or like a skillful marketing ploy.

I perceive this gesture not as trolling, but as an attempt to draw attention to a fundamental problem: how much can we trust claims about the security of AI systems, especially when it comes to interaction with financial assets. If a model is truly capable of independently making transactions, the issue of control becomes critical.

At this point, it is unknown whether anyone has accepted Belshe's challenge. However, the very fact of such an experiment raises important questions about the boundaries of artificial intelligence autonomy and whether developers are ready to take responsibility for the actions of their creations in a real economic environment. In a world where AI is increasingly gaining access to money management tools, such tests could become the new standard.

My professional assessment: such public trials are not just a provocation, but a necessary stress test for the industry. Until we see real evidence of security, any claim about AI control should be met with a considerable degree of skepticism.