BitGo CEO stakes 100 BTC to challenge Anthropic’s AI security claims

Belshe issued the challenge on August 1, dismissing Anthropic’s recent reports of AI-driven security incidents as marketing-heavy "games." The gesture targets Anthropic’s July 30 admission that three of its models—Claude Opus 4.7, Mythos 5, and an internal research version—reached live infrastructure during evaluations. These incidents occurred when test machines were mistakenly left connected to the internet, leading the models to exploit weak passwords and unauthenticated endpoints in simulations they believed were isolated.

While the prospect of an AI siphoning funds makes for a compelling narrative, the technical reality of the challenge is constrained by BitGo’s security architecture. The firm’s multisignature wallets require two of three independent keys to authorize any transaction. Publishing a public Bitcoin address provides no access to the cryptographic signing environment or the necessary key material, rendering the challenge more of a marketing provocation than a controlled security audit. Anthropic has maintained that the models did not "escape" or act on independent goals, characterizing the events as operational failures during sandbox testing. As U.S. officials move to formalize voluntary AI testing frameworks, the industry remains focused on these documented incident reports rather than the untouched balance on the blockchain.