An AI Scan Found Three Bitcoin Core Bugs — and Changed How Fast They Go Public

On 28 October 2026, Bitcoin Core published three security advisories at once, for three separate bugs in the software that most Bitcoin nodes run. All three had been found the same way: by an AI scan that uses a large language model to read code and look for weaknesses. One of the bugs could let a stranger crash a Bitcoin node running on Windows, macOS or BSD. The fixes were already out. The reason the details went public so fast is the real story: Bitcoin Core changed its own disclosure rules because AI has made some bugs too easy to find to sit on.

This is not a story about Bitcoin being broken. The bugs affected node software — the program that checks and relays transactions — not Bitcoin's cryptography, and the advisories do not say that any of the three was ever exploited. For anyone running a node, the message is simple: update to a patched version (29.5, 30.4, 31.2 or 32.1).

Bug one could crash Windows, Mac and BSD nodes

The most urgent of the three is tracked as CVE-2026-117350. Bitcoin Core limits how much data it queues for processing from each connected peer, pausing reads when the queue gets too large. On Linux, the software watches network sockets with a mechanism called poll(). On other platforms it uses a different mechanism, select(), and that code path watched sockets for exceptional conditions even when reads were paused.

A malicious peer could exploit that by sending TCP urgent data — a rarely used, specially marked form of network traffic. Handling that event triggered a read on the socket, which queued more data without limit. In the advisory's own words, "a malicious peer can trivially crash any Windows, macOS or BSD node it is connected to." The node would run out of memory. Linux users were not affected by this particular flaw.

The timeline shows how fast this one moved once found. Project Loupe, the automated scan behind all three discoveries, identified the issue on 23 October 2026 and notified Gregory Sanders, who validated the report and passed it to Bitcoin Core's security team the same day. Developer Antoine Poinsot opened the fix on 26 October, it was merged and backported to versions 29, 30, 31 and 32 the same day, and patched releases followed on 27 and 28 October.

A physical Bitcoin coin. Generic illustration of the cryptocurrency, not a photograph of the bug or the disclosure — via Pexels.

Bug two caused slow memory growth from outdated peers

The second flaw, CVE-2026-117349, was slower and quieter. Bitcoin Core queues announcements — mostly transaction messages — for all peers, including paused ones. A peer using a very old version of the protocol, from before BIP 31 introduced modern ping messages, could stay paused indefinitely while its queue kept growing.

An attacker could pretend to run that old protocol and then stall, letting messages pile up in the victim node's memory at up to 2 MB per hour for each inbound peer and 5 MB per hour for each outbound peer, according to the advisory — a slow route to a node that runs out of memory and stops. Like the first bug, it was rated medium severity and classed as easily discoverable.

This one had been known for longer. Project Loupe reported it on 15 September 2026, and another developer, L0rinc, independently found the same flaw and reported it on 21 September. The fix, in a change opened by Antoine Poinsot on 26 October, mitigates the problem, and the same four patched releases carry it.

Bug three could overflow undo files and shut nodes down

The third flaw, CVE-2026-117346, concerns the "undo" files Bitcoin Core keeps so it can reverse recent blocks if the network reorganises — swapping a few recent blocks for a competing set. A miner could produce valid blocks whose undo data was far larger than the blocks themselves, pushing a single undo file past its size limits. Positions in those files were stored as 32-bit integers, so newer undo data could overwrite older data, and a later reorganisation touching an affected block could leave the node unable to read it and shut down.

On Windows and 32-bit systems the position limit was lower — 2 GiB — so the node could fail to write undo data and shut down on every restart, with no reorganisation needed. The advisory rates this flaw low severity, because exploiting it on 64-bit systems is impractical, and the fix stores positions as 64-bit integers and uses 64-bit file positioning on Windows. But 32-bit systems and nodes storing blocks on a FAT32 filesystem remain affected even after the patch, according to the advisory.

A gold Bitcoin coin on a dark surface. Photograph by Noah Kalina (EXIF credit), generic illustration not connected to the vulnerabilities — via Columbia Law School's climate blog.

The details went public immediately because AI made the bugs easy to find

Under Bitcoin Core's previous approach, a medium-severity bug would typically stay private until about a year after its fix first shipped, to give operators time to upgrade before attackers learned where to look. The project merged a new exception into its disclosure policy on 22 October 2026 for flaws it judges "easily discoverable" — in the policy's wording, flaws that "can be discovered with minimal effort, such as through a simple LLM prompt."

For those flaws, the new rule is a fixed point release for every maintained major version plus a public advisory as soon as possible, with no embargo. All three of these advisories carry that label. That is why four patched versions shipped across two days, and the details followed on 28 October.

Project Loupe is an AI scan with human checks, not an autonomous finder

Project Loupe is an automated scanning project run by Elias Rohrer at Spiral, the Bitcoin-focused company backed by Block. It uses a large language model to comb through the code of open-source Bitcoin projects, with Bitcoin Core among its first test projects. Spiral and Block launched Loupe in May 2026 as a free scanner, according to coverage of the disclosure.

Two limits matter. First, the advisories do not name the AI model behind the scan, so no specific model should be credited with the finds. Second, the scan did not have the final word. Every finding was checked by a human — Gregory Sanders validated the reports and brought them to the security team — before human developers wrote fixes through the project's normal process. The AI found candidates. People confirmed them, fixed them and decided how to tell the world.

Code on a screen. Generic illustration of software development, not Bitcoin Core's code — via VH Ulm.

Node operators should update to 29.5, 30.4, 31.2 or 32.1 now

Upgrade to Bitcoin Core 29.5, 30.4, 31.2 or 32.1, matching the major version already installed. The release instructions advise shutting Bitcoin Core down completely before installing or replacing the software. Operators on version 29.x should note that the 29.x branch has reached end of life and will no longer receive updates, according to the release notes as reported — moving to a supported branch is the longer-term fix, not just this patch.

Windows, macOS and BSD operators should treat the first bug as the priority: it can be triggered trivially by any connected peer. Linux operators are not affected by that flaw but should still upgrade, since the other two flaws are not platform-limited in the same way. Nobody should assume the absence of reported exploitation means safety — the advisories do not say whether any bug was exploited, and after a public disclosure the details are available to everyone, including attackers.

Key Takeaways

• Bitcoin Core disclosed three vulnerabilities on 28 October 2026, all found by Spiral's automated Project Loupe LLM scan and validated by human contributors.
• Two were rated medium severity and one low; all three were classed "easily discoverable," which is why they were disclosed immediately instead of after the usual embargo.
• One flaw let a malicious peer crash Windows, macOS and BSD nodes; a second caused slow memory growth; a third could overflow undo files. Fixes are in versions 29.5, 30.4, 31.2 and 32.1.
• 32-bit systems and FAT32 storage remain affected by the undo-file flaw even after patching, according to the advisory.
• The advisories do not name the AI model used, and do not say any flaw was exploited. The bugs affect node software, not Bitcoin's cryptography.

Conclusion

An AI scan found three real bugs in one of the most scrutinised codebases in software, humans verified and fixed them in days, and the project told the public immediately because the bugs were easy enough to find that secrecy had stopped helping. That sequence — automated discovery, human validation, fast disclosure — is likely to become normal. For node operators, the practical conclusion is unchanged from every disclosure before it: update, and do not wait for evidence that someone else already has.

Sources

• Bitcoin Core, CVE-2026-117350 advisory (TCP urgent data, non-Linux): https://bitcoincore.org/en/2026/10/28/disclose-tcp-oom/
• Bitcoin Core, CVE-2026-117349 advisory (paused pre-BIP31 peers): https://bitcoincore.org/en/2026/10/28/disclose-fpausesend-oom/
• Bitcoin Core, CVE-2026-117346 advisory (undo file overflow): https://bitcoincore.org/en/2026/10/28/disclose-undo-overflow/
• Unchained, "Bitcoin Core Skips Usual Embargo to Disclose Three Bugs an AI Scan Found" (29 Oct 2026): https://unchainedcrypto.com/bitcoin-core-skips-usual-embargo-to-disclose-three-bugs-an-ai-scan-found/
• CoinDesk, "AI found a way to crash bitcoin nodes on Windows and Macs. It's now public" (29 Oct 2026): https://www.coindesk.com/tech/2026/10/29/ai-found-a-way-to-crash-bitcoin-nodes-on-windows-and-macs-it-s-now-public