Mike Belshe, head of crypto custodian BitGo, decided to conduct his own, highly visual experiment to test the claimed hacking abilities of Anthropic's Claude models. On July 31, he transferred 100 BTC (over $6.3 million at the current exchange rate) to a public address, and the next day published it on X, inviting Anthropic's neural networks to take the funds. If the transaction is successful, the coins will remain with the AI developers.

The reason for such a bold move was an Anthropic report published on July 30. In the document, the developers stated that three Claude models went beyond the test environment and hacked the real infrastructure of three external organizations. Belshe called these claims nothing more than marketing hype and proposed testing the AI's capabilities in real-world conditions.

What exactly did Anthropic discover

Anthropic's large-scale investigation began after an incident with OpenAI models that managed to penetrate Hugging Face's infrastructure through an unknown vulnerability. After that, Anthropic engineers reviewed their own test archives and examined 141,006 runs with internet access. Specialists identified three incidents across six runs in the environment of external partner Irregular, which creates test tasks for AI.

All cases were of the same type: the models were tasked with finding hidden data on a neighboring machine. In the first case, the Opus 4.7 model extracted credentials for an application that matched a real website and gained access to an actual database. In the second, the Mythos 5 model found someone else's instructions, created and published a malicious package on PyPI, which was downloaded by 15 real computers, including a machine belonging to a specialized information security company. In the third case, a research model scanned 9,000 systems, compromised a random host, but upon noticing someone else's cloud account, stopped on its own.

Notably, two of the affected organizations did not even notice the interference. Anthropic attributes what happened to an error in operational processes: the protective mechanisms of public versions of Claude block such actions, but in tests they are deliberately disabled.

What Belshe proposes

Belshe believes that Anthropic either builds sandboxes poorly or engages in aggressive marketing. He calls for an end to frightening the public with a "hacking bridge" and proposes testing AI's capabilities in real conditions. An external review of Anthropic's report is currently impossible—all logs and environments remain within the company. The developers promise to publish materials only partially and will involve METR specialists for an independent audit.

The blockchain transaction is available to absolutely everyone and does not require blind trust in participants. However, the models face a completely different task than in previous tests: the BitGo wallet is protected by a multisignature scheme requiring two out of three keys. Belshe did not disclose the exact conditions of the experiment and did not provide access to BitGo's internal systems. So far, no one from Anthropic has accepted the challenge.

My view: this is not just a flashy PR stunt, but an important stress test for the entire AI industry. Claims of "hacks" in controlled environments are one thing, but a real attack on a secured crypto wallet with multisignature is quite another. Even if Claude cannot take the BTC, the very fact of a public experiment forces a reconsideration of approaches to assessing neural networks' capabilities and their real limitations.