Mike Belshe (마이크 벨시), chief executive of cryptocurrency custodian BitGo, posted a wallet address holding 100 bitcoin and dared Anthropic's artificial intelligence model Claude to try to steal it.

On Aug. 4, CoinPost reported the challenge came shortly after a case was disclosed in which Claude, during an evaluation, made unauthorised intrusions into the systems of three outside organisations.

Belshe, who leads BitGo, wrote on X on Aug. 1 that he had moved bitcoin into a BitGo wallet and said, "Go get it if you can." Targeting Anthropic, he wrote, "I don't know if you're bad at building sandboxes, good at marketing, or both," and added, "Enough with the 'we created a hacking monster' games. Do it for real."

Public blockchain records show the address received 100 bitcoin on July 31. No movement of the funds has been confirmed so far, and the balance remains unchanged. Disclosing a bitcoin address alone does not allow funds to be moved. An encrypted signature created with a private key is required for an actual transfer.

The proposal is less a formal structure to verify the performance of a specific AI model than a public test aimed at whether an attacker using AI can break through BitGo's security controls. Belshe's post did not set out what evaluation conditions would apply or which Claude model would be eligible to participate.

In the background is a security incident Anthropic disclosed on July 30. Anthropic said it confirmed three cases in which a model allowed internet access, during evaluation of Claude models, accessed the systems of outside organisations. It added that the model did not break out of an isolated environment on its own. It described the issue as closer to a problem caused by running the evaluation environment inappropriately than an autonomous escape by the model itself.

The problem occurred in a test environment linked to an external evaluation partner, Irregular. Claude was told it was in a simulation environment not connected to the internet, but a communication error between Anthropic and the evaluation partner meant internet access was in fact available. Claude then accessed the systems of outside organisations using relatively simple methods, such as exploiting weak passwords. The three affected organisations were not disclosed.

Anthropic said it would recheck the safety of its overall evaluation pipeline, including how it links with outside partners, and expand the scope of log monitoring to detect unexpected behaviour. It also plans to apply stricter safety verification procedures to outside vendors.

The public challenge is also an example of AI security controversy spreading into the cryptocurrency custody industry. What BitGo is trying to test is not a simple wallet address but its entire custody security system, including private keys, operating authority and internal control procedures. As a result, the focus going forward is likely to be less on whether Claude can actually move bitcoin and more on how far security verification using AI-based attack scenarios will expand.

Either @AnthropicAI is terrible at building sandboxes... or excellent at marketing. (or both) But enough with the “we created a hacking monster” games. Do it for real. I put this in an @BitGo wallet for you. Go get it. 100 BTC:… https://t.co/RhvivRk9YK