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

What it is

In October 2026 Bitcoin Core maintainers merged a security policy change written by Antoine Poinsot of Chaincode Labs. It adds a triage flag for bugs "that can be found with a simple prompt to an LLM". Flagged bugs skip the long embargo.

How it works

Before the change, Core graded every reported vulnerability by severity. Low-severity flaws were disclosed two weeks after the next major version shipped. Medium and high flaws stayed under embargo until two weeks after the last affected version reached end of life.

flowchart TD
    A[Report triaged] --> B{Easy to find with an LLM prompt?}
    B -->|Yes| C[Backport to all supported versions]
    C --> D[Point release ASAP plus advisory with some details]
    B -->|No| E[Old severity-based embargo]

The policy calls the old embargo the wrong tradeoff for easy bugs, because it "achieves little but it delays the availability of a fixed release." Poinsot cites more reports of easily rediscovered bugs, which he attributes to LLMs lowering the bar for finding issues and for spotting covert fixes.

Preconditions / Affected

The rule applies only to bugs the triagers judge easy to rediscover. The page says no bug has been flagged under it yet. It does not say who makes the call beyond "during triage".

What to do

The page gives no commands. For anyone running Core nodes, the practical effect is that point releases with security advisories may arrive without warning and cover every supported version. Maintainers of other projects can compare their own embargo rules against this one.

Caveats

• Poinsot says fixes "won't have as much 'bake time'" in testing before reaching users.
• Antoine Riard wanted to keep the option of shipping a fix without an advisory and holding details back case by case. Poinsot refused, saying that alerting users without details could cause "a rushed migration by users who would otherwise not be affected."
• The context the piece gives: 16 volunteers calling themselves the Bitcoin Red Team ran AI models over 390 projects in about a day in late July, logging 4,962 findings, 85 critical. A BTCPay Server flaw they reported privately was exploited within days, draining merchants' Lightning nodes. Core Lightning told operators to disconnect after AI-generated reports proved real.
• The Coldcard firmware flaw, dormant since 2021, was not found by AI. It let attackers drain 594 BTC in 25 minutes on July 30, with losses near $114 million across four sweeps.