AI is finding real bugs in Bitcoin's payment software — and the panic travelled faster than the fix

The developers of Core Lightning, one of the main programs people run to send and route bitcoin over the Lightning Network, told node operators this week to install a forthcoming security update promptly after a wave of AI-generated bug reports turned up several genuine vulnerabilities. The team is holding the technical details private for about two weeks so that operators can patch before attackers can study what was fixed.

By the project's own account, the reports began arriving around the middle of August, from several sources, over roughly ten days. A small team plus outside contributors spent that time checking which submissions described real weaknesses. Several did. On August 23 the project said it would ship signed, reproducible binaries containing fixes, and that it would stop supporting earlier releases, including version 26.04. It has not said how many flaws are being fixed, what an attacker could do with them, or whether anyone has exploited them; the regularly scheduled 26.09 release is still planned for late September.

The advice itself then got mangled in transit. A screenshot circulating on Bitcoin social media urged operators to shut their nodes down immediately, and the developers had to push back: switching a machine off is the worse option. A Lightning channel works by two participants locking bitcoin into a shared balance and updating it between themselves, settling on the Bitcoin blockchain only when the channel closes. That means a node has to keep watching the chain in case the other side tries to close using an outdated balance. A running node can respond and defend the funds; a powered-off one cannot. Restarting with the --offline flag stops payments routing in or out while the software keeps watching.

The episode is part of a broader shift. The Bitcoin Red Team, a group of sixteen developers, used AI models to sweep 390 Bitcoin repositories in late July, producing nearly 5,000 findings, 85 of them initially rated critical, in about 27 hours of work. Volume is not the same as accuracy: Google revised its open-source bug bounty programme in March after a surge of AI-generated submissions containing incorrect information and invented exploit paths, and humans still have to validate every report. But some of the findings are real. Earlier this month a flaw in BTCPay Server exposed the credentials controlling Lightning nodes and attackers drained funds before a fix shipped, and in July the maker of the Coldcard wallet said an attacker appeared to have used AI to find a weakness in how its older software generated wallet seeds.

Most beginners do not run a Lightning node, and if you use a custodial app your provider handles this on your behalf. The pattern is still worth learning, because it is the shape scams imitate. Real fixes arrive as signed releases through a project's official channels, with a version number you can check on your own device. An urgent instruction that reaches you as a screenshot, a forwarded message or a direct message is exactly how people get talked into installing fake updates. When you hear that a wallet needs an emergency upgrade, go to the project's own site or store listing and verify the version yourself. This is information, not advice.