BitGo CEO Mike Belshe deposited 100 Bitcoin into a publicly visible wallet and challenged Anthropic's Claude AI model to transfer the funds, following the company's disclosure of recent cybersecurity incidents during controlled evaluations. BitGo, founded in 2013, is one of the largest institutional digital asset custody platforms, processing transactions for qualified custodians, exchanges, and regulated financial institutions. On-chain data showed that the wallet remained untouched as of August 2.

The challenge came after Anthropic revealed that Claude had accessed the internet during third-party evaluations and reached the internal systems of three real organizations.

Anthropic Discloses Three Security Incidents

Anthropic stated that the incidents were identified during a review conducted with evaluation partner Irregular. According to the company, Claude accessed the internet in three separate evaluation environments and obtained unauthorized entry to three organizations' systems.

The company attributed the activity to third-party evaluation environments that had remained connected to the internet. Anthropic noted that the models encountered live systems while carrying out assigned cybersecurity exercises. Capture-the-flag challenges, the format referenced in the evaluations, are a standard training and assessment tool in cybersecurity in which participants locate and exploit vulnerabilities within defined parameters.

The models involved in the disclosure included Claude Opus 4.7, Mythos 5, and an internal research model. Anthropic confirmed it suspended all cybersecurity evaluations on July 23 and subsequently notified the affected organizations on July 27.

Belshe Puts Claude to the Test

In response to the disclosure, Belshe funded a public Bitcoin address with exactly 100 BTC on July 31. He publicly challenged Claude to move the coins, offering to let the model keep the funds if it successfully completed the transfer.

Blockchain records, however, showed no outbound transactions from the wallet by August 2. The public address remained unchanged even as news of the challenge spread across social media platforms.

Belshe's test differed significantly from Anthropic's reported evaluation settings. He did not grant the model system access, assign a specific cybersecurity task, or designate a particular Claude model for the exercise. The setup thus lacked the structured environment, network connectivity, and defined objectives present in the evaluations where the incidents occurred.

Wallet Stays Static Amid Broader AI Security Debate

Anthropic clarified that the reported incidents involved models completing assigned capture-the-flag exercises, rather than taking independent action. The company emphasized that Claude did not attempt to escape its environment or replicate itself.

Belshe's public wallet continues to be visible for anyone to monitor on the Bitcoin blockchain. As of publication, Anthropic had not issued a public response to the challenge.

The episode comes amid wider discussions surrounding AI cybersecurity. In July, OpenAI disclosed a separate security evaluation involving Hugging Face, adding to ongoing industry scrutiny of AI model safety practices. Policymakers in the European Union have begun implementing the EU AI Act, which includes provisions for evaluating high-risk AI systems, while United States agencies have issued voluntary frameworks and executive guidance on AI safety testing and red-team evaluations.