Rob Hamilton uses Chinese AI models to audit Bitcoin
- A five-year-old Coldcard firmware bug let automated scripts drain over $114 million in Bitcoin.
- Marathon's private transaction relay protected 5,600 Bitcoin from front-running attackers during emergency key migrations.
- US AI usage limits forced security researchers to use cheap Chinese models like Kimi K3 to scan repositories.
AI scanners are ripping through open-source Bitcoin repositories faster than developers can patch them.
The crisis began when a 2021 Coldcard firmware update bypassed hardware random number generators, reverting devices to weak software entropy. Attackers automated address sweeps, stealing 1,816 Bitcoin across 5,200 addresses within seconds of receiving funds. On Ungovernable Misfits, researcher Orange Surf tracked over 5,600 Bitcoin moving through Marathon's Slipstream private transaction service to protect multisig migrations from front-running bots watching the public mempool.
In response, white-hat security researchers formed the Bitcoin Red Team to find flaws before attackers could exploit them. Led by Rob Hamilton and developer Cali, the initiative spent over $40,000 using models like OpenAI's GPT Sol and Anthropic's Claude Opus to scan 171,000 lines of code across 390 repositories. The automated harness flagged 85 critical vulnerabilities, forcing projects like Bolts exchange to halt operations to apply patches.
As defensive scanning expanded, researchers ran into administrative barriers with Western AI providers. On The Jack Mallers Show, Strike CEO Jack Mallers explained that US AI safety guardrails and rate limits hindered white-hat audits, forcing researchers and security teams to adopt open-weight Chinese models like Kimi K3 instead. Attackers operate without regulatory constraints, making restrictive access policies a distinct disadvantage for cyber defenders.
"The bugs were always there. AI simply made finding them virtually free."
- Jack Mallers, The Jack Mallers Show
The rapid vulnerability discovery also exposed deeper flaws in open-source software incentives. Coldcard’s firmware remained vulnerable for five years despite being publicly viewable on GitHub. Mallers noted on his show that Coldcard’s restrictive Commons Clause licensing discouraged third-party developers from spending commercial resources to audit the codebase, contrasting it with Bitcoin Core’s open MIT license that incentivizes institutional battle-testing.
Meanwhile, the public handling of the exploit severely damaged developer trust. On Stacker News Live, hosts Carr and Keon discussed how Coinkite founder NVK deleted historical social media posts while lead developer Doc Hex was revealed to have used a secondary online identity to approve his own code. Former community allies distanced themselves as users realized that informal social reputations had replaced formal engineering reviews.
The security fallout is accelerating an institutional pivot away from self-custody. Host David Bennett warned on Bitcoin And that the sudden wave of automated drains is driving self-custodial holders toward regulated custodians like Coinbase or spot ETFs. Flawed software accomplished what years of direct regulatory pressure failed to do: pushing sovereign Bitcoin assets straight into corporate vaults.
"Flawed code accomplished what regulatory crackdowns could not: driving sovereign Bitcoin back into regulated vaults."
- David Bennett, Bitcoin And | Bitcoin & Economic News
Code visibility without defensive AI deployment is now a liability.