BitGo Chief Executive Mike Belshe has turned a high-profile debate over artificial intelligence and digital asset security into a public stress test, funding a wallet with 100 BTC—worth roughly $6.3 million at the time—and challenging Anthropic’s Claude AI models to compromise the funds. The stunt spotlights growing industry questions about whether advanced AI systems can meaningfully threaten the security of modern crypto custody.
Public Wallet Challenge
Belshe funded a publicly known bitcoin wallet with 100 BTC and issued an open challenge to see whether AI systems could move the funds without authorization. While the wallet address is public by design, the private keys required to sign transactions are not. Without the private key—or a successful compromise of the systems that protect it—bitcoin cannot be spent.
AI Capabilities and Security Debate
The move followed a disclosure by Anthropic regarding the capabilities of its Claude models, which reignited discussion about AI’s potential to aid cyberattacks. Security researchers and crypto custodians have debated whether current AI systems can autonomously discover and exploit real-world vulnerabilities—such as weaknesses in key management, operational processes, or user behavior—versus merely assisting human attackers with known tactics.
In the context of Bitcoin, the underlying cryptography remains robust: knowledge of a public address does not reveal the corresponding private key. Any successful theft would therefore require an operational security failure, a device or infrastructure exploit, social engineering, or similar non-cryptographic vectors.
Implications for Crypto Custody
BitGo, an institutional digital asset custodian, and its peers employ layered defenses that commonly include hardware security modules, multi-party computation (MPC), and offline storage. Public challenges of this kind are intended to pressure-test assumptions and raise awareness about the difference between cryptographic security and the practical realities of securing keys and workflows in the wild.
The outcome of the challenge—regardless of success or failure—will likely inform ongoing discussions about threat modeling, red-teaming with AI tools, and best practices for safeguarding digital assets as AI systems continue to advance.