Bitcoin Core changes security rules as AI makes finding code bugs dangerously easy

Bitcoin Core, the software most of the bitcoin network runs on, will stop sitting on bug fixes for months when the bug is one an AI could find again, after its maintainers merged a change to the project’s security policy on Thursday.

Core previously graded every reported vulnerability by severity and timed disclosure accordingly. Low-severity flaws were revealed two weeks after the next major version shipped. Medium and high ones stayed under embargo until two weeks after the last affected version reached end of life.

The logic was that shipping a fix quietly and explaining it much later kept attackers from learning about a hole before the people running nodes, the computers that check and relay every bitcoin transaction, had upgraded.

But the old rule was built for a world where finding a bug took skill, and the spread of AI tools has broken how that process worked.

“We recently experienced an increase in the number of reports for security vulnerabilities that can be easily (re-)discovered,” Antoine Poinsot, the Chaincode Labs developer who wrote the change, said in his proposal.

“This is largely due to LLMs drastically reducing the bar to find certain classes of issues or identify covert fixes.”

How it works

Core can now flag a bug as easy to discover during triage, which in practice means one that “can be found with a simple prompt to an LLM.” Those bugs get a fix copied back to every version Core still supports, shipped in a point release “as soon as possible” along with a security advisory carrying some details.

“For these issues that are easy to find, this is the wrong tradeoff,” the policy reads of the old embargo. “The embargo period achieves little but it delays the availability of a fixed release.”

Fixes “won’t have as much ‘bake time’” in testing before they reach users, however, Poinsot wrote.

Not every developer agreed on how far to go. Antoine Riard, a longtime bitcoin security researcher, wanted Core to keep the option of shipping a fix without an advisory and holding the details back case by case.

Poinsot refused. “Alerting users of a security vulnerability without any details could create more risk unnecessarily,” he wrote, pointing to “a rushed migration by users who would otherwise not be affected.”

A volunteer group of 16 developers calling itself the Bitcoin Red Team ran AI models across 390 bitcoin projects in about a day in late July and logged 4,962 findings, with 85 of them critical.

Within days a BTCPay Server flaw the group had reported privately was being exploited on live merchant servers. The attacks drained the Lightning nodes merchants use to take small bitcoin payments. Core Lightning, one of the main Lightning implementations, told operators to disconnect after AI-generated reports to its small team turned out to be real.

A bigger loss came from a bug no AI needed to find. A flaw in Coldcard hardware-wallet firmware, dormant since 2021, let attackers drain 594 BTC in 25 minutes on July 30, and losses neared $114 million across four sweeps. When Coldcard shipped its fix, it said AI had helped catch more bugs along the way.

No bug has been flagged under the new rule yet, but a secret any model can regenerate in one prompt was never going to protect users for long.