Mike Belshe, co-founder and chief executive of BitGo, a crypto custodian, dared Anthropic's Claude to take 100 BTC from what he described as a company wallet on Aug. 1, turning the AI developer's security disclosure into a roughly $6.3 million public challenge.
"Either Anthropic is terrible at building sandboxes... or excellent at marketing. (or both)," Belshe wrote on X. "But enough with the 'we created a hacking monster' games. Do it for real."
- The Coinformer verified that the published address received exactly 100 BTC in one transaction at 5:07 pm UTC on July 31, about 22 hours before Belshe's post.
No door to kick down
The challenge comes shortly after Anthropic said its evaluation partner accidentally left internet access open, allowing Claude models to reach real organizations while attempting simulated hacking exercises.
The models used weak passwords, unprotected endpoints, credentials exposed through a debug page and SQL injection. They didn't defeat encryption or deliberately escape through an unknown vulnerability.
Anthropic found six affected runs across three incidents.
- The incident wasn't unique to Anthropic. In late July, OpenAI said GPT‑5.6 Sol, one of its AI models, and an internal prototype exploited a previously unknown flaw in Artifactory, a software package proxy, to escape an isolated evaluation network and access Hugging Face's production systems.
- Unlike Claude, the OpenAI models found a genuine zero-day, although they too were running without normal production safeguards.