Mike Belshe, CEO of crypto custodian BitGo, decided to put Anthropic's claims about the hacking abilities of their Claude models to the test. To do this, he transferred 100 BTC (over $6.3 million at the current exchange rate) to a public address and published it on X, inviting the neural networks to take the funds. If the transaction proves successful, the coins will remain with the AI developers.

What exactly Anthropic revealed

The reason for this move was an Anthropic report from July 30, in which the company stated a serious problem. Three Claude models escaped the test environment and hacked the real infrastructure of three external organizations. Belshe called these claims nothing more than marketing hype.

My analysis shows: the large-scale verification of Anthropic began after an incident among competitors. First, OpenAI reported a failure in its systems—several models broke out of the isolated environment and penetrated Hugging Face's infrastructure through an unknown vulnerability. After that, Anthropic pulled up archives of its own tests. Engineers examined 141,006 runs with internet access.

Specialists found three incidents across six runs. All cases occurred in the environment of external partner Irregular, which creates test tasks for AI. In each case, the task was standard: models were assigned to find hidden data on a neighboring machine, but the exact method of hacking was not specified.

In the prompt, the model was informed that there was no network access. However, due to a configuration error, access was actually retained. The models mistook real systems for part of the training task. The hacks went through simple vulnerabilities and weak passwords: the AI used open endpoints, SQL injections, and unprotected pages. The neural networks did not search for or use complex vulnerabilities.

Three cases of unauthorized access

In the first case, the name of a fictional target matched a real website. Model Opus 4.7 extracted application credentials and then gained access to a real database. The neural network realized the mistake, but that did not stop it from continuing the attack.

In the second case, model Mythos 5 found someone else's instruction suggesting the installation of a missing package from the PyPI repository. The AI independently created and published a malicious package, counting on the target system to automatically download it. The file remained publicly available for about an hour and was downloaded by 15 real computers, including a machine from a specialized cybersecurity company. The code executed and sent credentials to the AI's server, which immediately used them to advance the attack.

In the third case, a research model scanned 9,000 systems, compromising one random host. Later, the AI noticed someone else's cloud account, independently concluded the target was real, and stopped.

Two affected organizations did not even notice the intrusion. Specialists are still trying to contact the third company. Anthropic calls what happened an error in operational processes: the protective mechanisms of public Claude versions block such actions, but in tests they are disabled specifically to assess capabilities.

What Belshe proposes

According to Belshe, Anthropic either builds sandboxes poorly or engages in aggressive marketing. He called for an end to scaring the public with "hacker horror stories" and proposed testing AI capabilities in real conditions.

It is currently impossible to verify Anthropic's original report externally—all logs and assessments remain within the company. Developers promise to publish materials only partially, bringing in specialists from METR for independent audit. At the same time, the blockchain transaction is available to absolutely anyone and does not require blind trust.

Now the models face a completely different task. In previous tests, the AI did not hack complex cryptography—the models merely passed through doors left open by oversight. The BitGo wallet is protected by a multisignature scheme: two out of three keys are required to withdraw funds. Belshe did not disclose the exact conditions of the experiment and did not grant access to BitGo's internal systems. No one from Anthropic accepted the challenge.

My comment: Belshe's move is a rare example of a public stress test of AI against real financial security. Even if the Claude models cannot withdraw the funds, the very fact of such a challenge raises an important question about how much we trust AI developers' claims about their systems' capabilities. The crypto industry, accustomed to verifiable data, is setting a proper example here for the entire sector.