Bitcoin's Defenders vs the AI Guardrails
Bitcoin security researchers say restricted cybersecurity models from OpenAI and Anthropic refuse vulnerability-discovery work — AnchorWatch's Rob Hamilton described being cut off 19 minutes into a session — pushing the volunteer Red Team toward open-weights Chinese models like Moonshot's Kimi K3, per Decrypt and Bitcoin Magazine. Industry figures are now calling for formal partnerships with AI labs for authorized defensive access, citing incidents including Boltz's shutdown under AI-assisted attacks.
Red Team: the guardrail protects the attacker
Rob Hamilton, who runs a Bitcoin insurance company, got 19 minutes into hunting for flaws in Bitcoin infrastructure with OpenAI's cyber model before it shut him down. He's now using open Chinese models like Kimi K3 instead, and hates that he has to.
The safety rule that stops an AI helping someone find a money-stealing bug only stops the honest researcher who plays by the rules. The thief scraping wallet code doesn't ask permission.
So the guardrail meant to prevent harm ends up shielding the attacker and blinding the defender. On code holding a trillion dollars, where one bug drains someone's savings, that's not caution. It's an own goal.
Does 'safety' mean anything if it only ever binds the good guys?
4 sources
- Bitcoin Companies Want Help From AI Labs to Guard Against Hackers · decrypt.co · T2
- Non-Custodial Bitcoin Bridge Boltz Shuts Down After AI-Assisted Attacks · cryptopotato.com · T2
- 'Bitcoin Is Burning': Red Team Turns to Chinese AI to Find Flaws · decrypt.co · T2
- Chinese AI Beats Restricted OpenAI and Anthropic Cybersecurity Models, Bitcoin Industry Warns · bitcoinmagazine.com · T1