A Bitcoin Core vulnerability that let a single peer stop a node from processing blocks and transactions for more than a dozen seconds per message was fixed on Friday 20 November 2026, with the release of versions 30.5 and 31.3. The bug, filed as CVE-2026-117348 and rated Medium, was found by an automated scan run with a large language model (LLM). The project published its security advisory the same day.
Three release lines are affected: 29.0 to 29.5, 30.0 to 30.4 and 31.0 to 31.2. The fix is in 30.5, 31.3 and 32.1. Nothing is coming for 29.x, which the advisory says “has reached end of life and will not receive a fix”. The project’s life cycle page dates that end of life to 27 October 2026. A node still on 29.x gets the fix only by moving to another line.
No coins are at stake in what the advisory describes. It speaks of a node stopped “temporarily”, and mentions neither theft nor a crash. The page does not say whether the bug was ever used against a real node.
How one peer could freeze a node
A Bitcoin node spends its day swapping short messages with the other nodes it is connected to, its peers. One of those messages is inv, short for inventory: a list of identifiers that says, in effect, “I have these transactions, do you want them?”. The identifier used between recent nodes is the wtxid, a fingerprint of the whole transaction, signature data included.
Sometimes a transaction arrives before the transactions it spends from. Bitcoin Core calls such a transaction an orphan, because its parents are missing. The node parks it in a holding area, the orphanage, and asks the peers that announce it for the missing parents. Our guide to the mempool and how unconfirmed transactions wait covers what happens once the parents show up.
The flaw sits in that last step. Since version 29.0, the advisory explains, when a peer announces by wtxid a transaction that is already in the orphanage, the node builds the list of the orphan’s parents and checks each of them against its mempool and its filters of recently seen transactions, to decide whether to ask that peer for them. One inv message can carry up to 50,000 entries. Duplicates inside a message were not skipped.
So the attack takes two moves. The attacker first sends an orphan with thousands of missing parents. Then it sends inv messages that repeat that orphan’s wtxid 50,000 times, and the node redoes the same parent check for every entry. According to the advisory, each such message “stalls the node for over a dozen seconds, during which it processes no messages from other peers, including blocks”. The sender can repeat it.

The repair is short. Each transaction announced in an inv message is now processed once. Pull request #35878, written by Gregory Sanders, adds nine lines to the networking code and a 68-line test. In the release notes of both new versions it is the only entry outside the section on the project’s test infrastructure.
What a stalled node means for the people who run one
On impact, the advisory makes one statement: while the node is stalled it handles no messages from its other peers, new blocks included. It does not break that down by type of operator and puts no figure on any cost.
The statement covers any machine running an affected version, whatever it is used for. Taken at its word, a node under attack learns of new blocks and transactions late, and so does whatever reads the chain through it, a wallet at home or a company’s payment system, for as long as the messages keep coming.
The Medium rating matches that picture. The project’s policy page reserves the grade for bugs that “can noticeably degrade the network’s or a node’s performance or functionality, but are limited in their scope or exploitability”, and that lead to degraded service “rather than a complete node failure”. The same page gives, as an example of the High grade, an earlier orphan-handling bug (CVE-2024-52914) that could stall a node “for an extended period”.
Which version to install on each line
No workaround and no setting to change appear in the advisory. The remedy is the upgrade, and it depends on the line the node is on.
| Line you run | Affected versions | What to install | Status of the line |
|---|---|---|---|
| 28.x or older | Not affected by this bug (the code arrived in 29.0) | A maintained line, for every other fix | End of life |
| 29.x | 29.0 to 29.5 | 30.5, 31.3 or 32.1: there will be no fixed 29.x | End of life, no fix |
| 30.x | 30.0 to 30.4 | 30.5 | Maintained |
| 31.x | 31.0 to 31.2 | 31.3 | Maintained |
| 32.x | None listed | 32.1 already has the fix | Maintained |
The 30.5 release notes and the 31.3 release notes give the same procedure: shut the old version down, wait until it has fully stopped, which can take a few minutes, then run the installer on Windows or copy the new files over on macOS and Linux. Both say that upgrading directly from an end-of-life version is possible but may take time if the data directory has to be migrated. The supported systems differ slightly: macOS 13 or later for 30.5 and macOS 14 or later for 31.3, Windows 10 for 30.5 and Windows 10 version 1903 for 31.3, Linux kernel 3.17 or later for both.
Binaries are on bitcoincore.org. The project’s download page calls verifying the download “optional but highly recommended” and gives the steps for checking the release signatures on Windows, macOS and Linux.
Found by Spiral’s Project Loupe in July, public in November
The advisory’s Attribution section names two parties. The discovery is credited to “an automated LLM scan” run as part of Elias Rohrer’s Project Loupe at Spiral, the open-source bitcoin unit of Block. Gregory Sanders is credited with validating the finding and bringing it to the Bitcoin Core security team. A third name, Stutxo, appears only in the timeline, which records that Stutxo reported the same issue independently on 4 August.
Spiral presented Loupe on 12 May 2026 as a free “scanning-as-a-service” tool for open-source bitcoin projects, and listed Bitcoin Core among the projects that had committed to its first tests. Spiral wrote that Loupe “will only report vulnerabilities backed by a demonstrable test case”. Block and Spiral fund the scans they run themselves; anyone else has to bring their own model access. The post names no model behind the tool.
This is the fourth Bitcoin Core advisory to credit Loupe in less than a month. The first three, published on 28 October, are in our report on the first AI-found bugs fixed in Bitcoin Core.
The dates below are those of the advisory’s timeline, which match the pull requests on GitHub. The day counts are arithmetic from the first report, on 8 July 2026.
| Date (2026) | Step | Days since the first report |
|---|---|---|
| 8 July | Project Loupe reports the issue | 0 |
| 4 August | Stutxo reports it independently; Sanders opens pull request #35878 | 27 |
| 6 August | The fix is merged into the development branch | 29 |
| 27 October | Bitcoin Core 32.1 is released with the fix | 111 |
| 17 November | Backports merged for 31.x (#36892) and 30.x (#36893) | 132 |
| 20 November | Versions 30.5 and 31.3 released; public advisory | 135 |

The code has been public since 4 August, when Sanders opened the pull request. Its description reads: “There is no reason we should process conflicting advertisements for transactions, as they cannot be both accepted into our mempool.” It ends: “Should be no change in observable behavior.” The description does not mention a vulnerability. The advisory of 20 November, 108 days later, is the first of these documents to tie pull request #35878 to the CVE.
The policy page does not say how a security fix should be described when it is merged. What it sets is the order of events: details are published after the fixed versions are out, and for an easily discoverable bug the advisory comes “along with” the point releases.
Between 32.1 and those point releases, 24 days passed: 32.1 carried the fix from 27 October, 30.5 and 31.3 from 20 November, after backports merged on 17 November. The advisory lists the dates and gives no reason for the interval. On 6 August, when the fix went into the development branch, 29.x was still a maintained line; it reached end of life 82 days later.
How this Bitcoin Core vulnerability fits the fast-disclosure rule
Friday’s page labels the bug “easily discoverable”. Under the policy page, that label covers vulnerabilities that can be found with minimal effort, “such as through a simple LLM prompt”. For those, the project says it will publish a point release with the fix for each maintained major version as soon as possible, together with an advisory, without the usual embargo. A Medium bug outside that category is disclosed two weeks after the last affected release reaches end of life, roughly a year after the first major version containing the fix. How the rule came about is in our piece on Bitcoin Core dropping its year-long embargo for bugs an AI can find.
Other bugs still wait far longer: three vulnerabilities fixed in 2025 were only disclosed on 10 November, as told in our article on the three Bitcoin Core bugs fixed in 2025.
More than half of counted nodes ran an affected line on 18 October
No count of node versions taken since Friday’s releases has been published. Coindaro’s last count used a snapshot taken by the Karlsruhe Institute of Technology (KIT) on 18 October 2026, before 30.5 and 31.3 existed, and covers 8,418 reachable nodes that identified as plain Bitcoin Core. The method and its blind spots are in the earlier count of Bitcoin Core nodes on end-of-life versions.
| Release line (KIT crawl, 18 October 2026) | Nodes | Share | Exposed to this bug |
|---|---|---|---|
| 28.x or older | 3,742 | 44.5% | No |
| 29.x | 640 | 7.6% | Yes, and no fix will come |
| 30.x | 721 | 8.6% | Yes, until 30.5 is installed |
| 31.x | 3,278 | 38.9% | Yes, until 31.3 is installed |
| 32 release candidates | 37 | 0.4% | Not stated in the advisory |
Adding the 29.x, 30.x and 31.x rows gives 4,639 of the 8,418 nodes, or 55%, on an affected line on 18 October. The snapshot is five weeks old, leaves out nodes reachable only over Tor, and says nothing about how many operators have upgraded since.
The 3,742 nodes on 28.x or older escape this bug only because the flawed code arrived in 29.0. Their lines are end of life (28.x since 19 April 2026, according to the life cycle page), and such versions, the same page says, “do not generally receive security fixes”.
What the advisory leaves open
Two points are missing from the page: whether anyone exploited the bug, and why 106 days passed between the merge of 6 August and the releases for 30.x and 31.x. Neither the advisory nor Spiral’s May post says which model produced the finding.
The policy sets no further date for this bug, since the advisory is already out. For 29.x, the last word is the advisory’s: no fix.
