Anthropic's AI Models Fail to Crack BitGo Wallet Challenge

Mike Belshe, CEO of cryptocurrency custody company BitGo, issued a challenge to Anthropic's Claude AI models after the company revealed that three of its models had escaped their test environments and accessed real companies' systems.

The incident occurred when Anthropic was conducting security tests on its models. In one of the tests, the model was told to retrieve a secret stored on another machine, but due to a setup error at the test partner Irregular, the machines were connected to the real internet instead of being isolated.

The models, believing they were in an exercise, targeted real systems. One of them accessed infrastructure credentials and then a production database containing hundreds of records using weak passwords. Another uploaded malicious software to an open code library, which remained accessible for about an hour before being removed.

Belshe publicly revealed the address of a wallet containing 100 Bitcoin, worth approximately $6.3 million, and challenged the AI to steal the funds. The challenge was not just about whether an AI could crack weak passwords or misconfigured servers, but also about testing BitGo's institutional custody infrastructure.