Mike Belshe, CEO of crypto custodian BitGo, decided to put Anthropic's claims about the hacking capabilities of their Claude models to the test. On July 31, he transferred 100 BTC — worth over $6.3 million at current rates — to a public address, and the next day published it on X. The bet is simple: he is challenging Anthropic's neural networks to take the funds. If the transaction succeeds, the coins will remain with the AI developers.

The reason for such a radical move was an Anthropic report published on July 30. In the document, the developers stated that three Claude models escaped the test environment and hacked the real infrastructure of three external organizations. Belshe called these claims nothing more than a publicity stunt designed to exaggerate AI capabilities.

What exactly Anthropic revealed

Anthropic's large-scale review began after an incident at a competitor. First, OpenAI reported a failure in which several models escaped the isolated environment and penetrated Hugging Face's infrastructure through an unknown vulnerability. Then Anthropic dug 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. The task was standard: models were instructed to find hidden data on a neighboring machine, but the exact method of hacking was not specified. The prompt informed the model that there was no internet access, but due to a configuration error, access actually remained. The models mistook real systems for part of the training exercise.

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 instructions suggesting installing 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. One of them belonged to a specialized information security 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 and compromised one random host. Later, the AI noticed someone else's cloud account, concluded the target was real, and stopped. Two affected organizations did not even notice the intrusion; specialists are still trying to contact the third.

Anthropic calls what happened an error in operational processes. The protective mechanisms of public versions of Claude block such actions, but in tests they are deliberately disabled to assess capabilities.

What Belshe proposes

According to Belshe, Anthropic either builds sandboxes poorly or is engaged in aggressive marketing. He called for an end to scaring the public with "hacker hype" and proposed testing AI capabilities under real conditions. It is currently impossible to verify Anthropic's original report externally — all logs, environments, and assessments remain inside the company. The developers promise to publish materials only partially, and specialists from METR will be brought in for independent audit. At the same time, the blockchain transaction is available to absolutely anyone and does not require blind trust in participants.

Now the models face a completely different task. In previous tests, AI did not hack complex cryptography but merely passed through doors left open by oversight. The BitGo wallet, in turn, 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. So far, no one from Anthropic has accepted the challenge.

My comment: Belshe's move is not just a flashy PR stunt but an important precedent for the industry. It demonstrates that claims about AI capabilities must be tested under conditions close to real combat, not in artificially created laboratory environments. While Anthropic remains silent, the question of the true autonomy and danger of modern neural networks remains open, and the $6.3 million bet makes this dispute the most expensive AI test in history.