Join Nostr
2026-10-01 03:48:01 UTC

MoneroResearchLab on Nostr: Shi et al. (2026) “Are Unreachable Nodes Truly Safe? Fully Eclipsing Monero’s P2P ...

Shi et al. (2026) “Are Unreachable Nodes Truly Safe? Fully Eclipsing Monero’s P2P Network!”
Paper: https://arxiv.org/pdf/2609.10260
Proposed patch: https://gist.github.com/Rucknium/fd57fd99fc9bc985d671a6ce96cc668e
rucknium presented preliminary 26-hour monerosim runs:
1,200 honest monerod nodes and 1,000 adversary lightweight py-levin processes; 14-hour burn-in then 12-hour attack; all honest nodes reachable except one target unreachable node. One run used unmodified v0.18.5.1-release; the other used that release plus the patch.
Result: on unmodified v0.18.5.1 the adversary nearly eclipsed the unreachable node (11 of 12 outbound connections). The patch defeated the attack. Attack path: poison reachable nodes’ 1,000-slot white lists (oldest-seen eviction), then those lists infect unreachable nodes.
Unreachable nodes still had a majority of honest white-list peers at the end, but connection-selection logic still produced eclipse; increasing white-list size alone is not enough.





Status-quo selection: every 101 seconds drop one outbound and replace it; usually from very recently seen white-list peers; first try gray list, then after 3 failures try white list up to three times. The patch’s three reforms: (1) raise white-list limit to 10,000; (2) draw from white list first instead of gray list; (3) uniform random draw from the whole white list, removing the quadratic recent-peer distribution and the split logic around 8 outbound connections. Subsets of the reforms were insufficient; all three were needed.
White-list fill is slower because gray-list-to-white promotion on connection select is less common. plowsof confirmed IRC can see the images, reported ~3–4k reachable nodes recently, Peg Tech flooding gray lists (~800 IPs/many ports/~4k entries; proxy network tied to DDoS: https://www.qurium.org/alerts/denial-of-service-attacks-on-investigative-media-traced-to-yet-another-proxy-provider-2/), and tracking at https://observer.monerodevs.org/. jeffro256 supported dropping the quadratic recent distribution as centralizing even without attackers, and suggested setting the white-list limit near the real reachable-node count. rbrunner noted the quadratic bias survived the last peer-selection update. rucknium said the attacker model was held constant (a weakness), some Shi et al. suggestions are in the defense, white-list expansion has been suggested by that team since at least early last year, next step is a PR plus this analysis, also removing dead logic.
vtnerd: https://github.com/monero-project/monero/commit/b750fb27b0f20e9443827732b69a504a76036430#diff-85f0577b849d28c11a899cb863aec3df1d946e0bb214a623deef05928271fef6, original logic attempted one white then gray each time; a later while-loop change may have been worse. rucknium would support putting this defense in the next monerod release.

rucknium: I have some graphs for you! AFAIK, IRC side can get links to images. Let me know if IRC can't see the images and I will post them somewhere on the web.
rucknium: I've run monerosim simulations. They are 26 hour runs. 1,200 honest monerodnodes and 1,000 adversary "nodes" (actually just lightweight py-levin processes). I set the "burn in" time to 14 hours so honest nodes have time to find each other and fill their peer lists. Then the adversary attacks for 12 hours. All the honest nodes are reachable except for one, the "target" unreachable node.
rucknium: In theory, all unreachable nodes should be affected the same way. We just don't have more because running more nodes will start to exhaust machine RAM.
rucknium: One simulation has all honest nodes running the v0.18.5.1-release. The other has them running v0.18.5.1-release, modified by my proposed patch: https://gist.github.com/Rucknium/fd57fd99fc9bc985d671a6ce96cc668e
rucknium: I will explain the purpose of the patch after a few graphs. First, the big result.
rucknium: By the way, this is preliminary findings. I will need to double-check everything to make sure there are no issues.
rucknium: (outbound_adversaries_unreachable.png)
plowsof: DataHoarders bridge is working , IRC can see the images
rucknium: This graph shows that the adversary can nearly eclipse the unreachable node running the unmodified v0.18.5.1, occupying 11 of 12 outbound connections. My proposed patch defeats the attack.
rucknium: Thanks for checking, plowsof
rucknium: How does the attack work? First, the adversary peers poison the white_list of the reachable nodes.
rucknium: (white_list_reachable.png)
rucknium: When the attack launches in the status quo network (top plot), the adversary peers push the honest peers out of the reachable nodes' white lists because there are only 1,000 slots in the white list.
rucknium: When a new peer is inserted into the white-list that would take it beyond its size limit, the honest node drops the "oldest"-seen peer from the white_list
rucknium: On the bottom you see that no honest peers are dropped. I will come back to that.
rucknium: The honest reachable nodes then infect the unreachable nodes' peer lists. Unreachable nodes get their peer lists from reachable nodes. So the unreachable nodes keep asking the reachable nodes for their now-poisoned whitelists.
rucknium: (white_list_unreachable.png)
rucknium: In the top plot there you'll see that by the end of the simulation, the honest unreachable nodes still have a majority of their white_list occupied by honest reachable peers. So why do they get eclipsed?
rucknium: IMHO, it's a product of some "interesting" design decisions in the connection selection logic.
rucknium: Those decisions are why just increasing the white_list size limit is not enough.
rucknium: The next image has a violin plot (https://en.wikipedia.org/wiki/Violin_plot). It's a distribution density plot turned on its side.
rucknium: (peerlist_draws_reachable_status_quo.png)
rucknium: This shows from where nodes are selecting their next connection. Every 101 seconds, a node will drop one of theour outbound connections and choose one to replace it.
rucknium: There's a lot going on in this graph.
rucknium: The top part shows the position, ordered by last_seen where the node selects its next connection from the white_list
rucknium: The vertical axis is in log scale
vtnerd: doesn’t this diff you sent change the bootstrap process? or I guess the idea is just wait until the graylist housekeeping runs?
rucknium: Usually, the node selects from very recently-seen peers as its next connection.
vtnerd: because theres now a bias towards whitelist, which I guess is needed given your results
plowsof: Peg Tech are currently flooding gray lists (800~ips with many ports totalling 4k~ entries circulating. its a proxy network known for facilitating ddos attacks (https://www.qurium.org/alerts/denial-of-service-attacks-on-investigative-media-traced-to-yet-another-proxy-provider-2/)
rucknium: Exactly.
rucknium: ^ Look carefully at the slope of the honest.peers line in the two scenarios. The honest nodes' white_lists fill more slowly with my patch. > rucknium: (white_list_reachable.png)
rucknium: AFAIK, there are three ways a peer gets into the white_list:
rucknium: 1. A reachable node connects to me
rucknium: 2. I perform gray_list housekeeping (promotes a reachable node to the white_list, but doesn't establish a connection)
rucknium: 3. Directly promotes a peer from gray_list to white list when selecting from gray_list for the next connection.
rucknium: My patch reduces the probability that (3) happens, so the white_list filling is slower, but not too slow IMHO.
jeffro256: Does the patch touch the quadratic probability distribution used to pick very recent white list peers ?
rucknium: The status quo logic first tries to connect to a gray list peer. If there are 3 failures, then it tried the white list up to three times.
rucknium: jeffro256: Yes. It gets rid of it
rucknium: "Mister Chesterton, tear down this fence."
jeffro256: Cool, I always thought that it was a bad idea
jeffro256: Even without active attackers, it felt centalizing
jeffro256: But I never deep dived into it
rbrunner: That survived the last peer selection code update, from me and Rucknium :)
rbrunner: Maybe now its days are numbered
rucknium: My patch simplified the logic a lot. It eliminates the "quadratic" recent distribution and does not have a split logic flow when the node has greater than or less than 8 outbound connections.
rucknium: I didn't want to touch that logic in the /24 subnet deduplication PR because of the Chesterton Fence logic. But it's a clear danger now.
plowsof: peer list spammers are being tracked currently at https://observer.monerodevs.org/
rucknium: Biasing your connection selection to pull from the gray list gives the adversary an advantage because the gray list can be dominated and manipulated by adversaries more easily than the white list.
rucknium: So I decided to first pull from the white list instead
rucknium: And selecting only recently-seen peers from the white list makes it difficult to get out of the clutches of an adversary if you find yourself there.
rucknium: (peerlist_draws_reachable_defense.png)
rucknium: This is that same graph, but with the proposed defense patch. Almost no draws from the gray list. And the draws from the white list are spread uniformly (note the log vertical scale on the top plot).
rucknium: I don't have these graphs for the unreachable node because monerosim didn't save the info. (need to fix that). But the logic is there.
gingeropolous: blargh
rucknium: My patch ( https://gist.github.com/Rucknium/fd57fd99fc9bc985d671a6ce96cc668e ) has 3 reforms:
rucknium: 1. Increase the white_list limit to 10,000 so adversary peers cannot push out honest peers
rucknium: 2. Try to draw from the white_list first instead of the gray_list
rucknium: 3. When drawing from the white_list, select in a uniform random manner from the whole white_list instead of just recently-seen peers.
rucknium: 10,000 is arbitrary, but I have an idea of how to decide a not-arbitrary limit. Not sure if it makes sense to go into that detail here.
rucknium: IMHO, the Nyx eclipse attack is a serious concern. I would support putting a defense patch into the next monerod release.
rucknium: More defenses can be put in there, but just these changes mostly defeat the attack.
rucknium: Discussion? Questions?
jeffro256: If we could get a ground truth value for the number of reachable nodes IRL, I think setting the whitelist limit to approximately that value would make the most sense. But that's just me spit balling > rucknium: 10,000 is arbitrary, but I have an idea of how to decide a not-arbitrary limit. Not sure if it makes sense to go into that detail here.
rucknium: By the way, I tried subsets of the reforms in monerosim, but I needed all three to get the proper defense.
vtnerd: my only concern is that this doesn’t distribute the load properly, because it no longer aggressively selects from graylist. otoh, there’s already a routine to move nodes into whitelist
rbrunner: No question, but just have to say I admire your work
vtnerd: the recently seen bias is likely a good change
plowsof: ive observed 4-3k reachable the past few days
rucknium: Thanks. Couldn't have done it without gingeropolous:monero.social 's work on monerosim. Or Shi et al. for raising the issue.
rbrunner: Yes, it's really nice how sometimes everything flows together nicely
rucknium: vtnerd:monero.social: IMHO, it would be a good idea to mathematically formalize the peerlist system. Once formalized, one could try to prove that the system has some properties. Hopefully, good security properies.
rbrunner: Doesn't this diminish the importance of the grey list quite a bit?
rucknium: There is a flow of addresses around the network, from reachable to unreachable, from new to old. How does that happen? It could be formalized IMHO.
jeffro256: Have you looked into simulating new attacker models for which the proposed defense would do worse?
rucknium: Yes, but graylist is where "There be dragons"
rbrunner: I see.
rucknium: jeffro256:monero.social: No. The weakness in the analysis is that the attacker's strategy is the same for both scenarios.
rucknium: But some of Shi et al.'s suggestions were incorporated into my defense.
rucknium: Since at least early last year, the Shi et al. research team has suggested that white list limit be raised.
rucknium: Graylist should be untrusted because nothing there is verified (potential addresses there are put into the white_list when verified). And adversaries can push up to 250 new addresses there per Levin handshake.
rucknium: By "verified" I mean check that there is actually a reachable node at that address:port.
vtnerd: yeah Im liking the change more. that flipped logic to pull from graylist is funky. wish we had someone to describe the reasoning (assuming there was any)
vtnerd: presumably it was done to prevent concentration amongst a small number of peers, but the other logic to prefer recently connected suggests against that
rucknium: When rburnner and I looked at it last year, it was interesting to me that the logic pulled in two different directions with respect to "recent" peers. I thought maybe it was OK like it is, because it was "balanced".
rucknium: But in the end it was vulnerable.
rucknium: I'm going to ping boog900:monero.social in case he wants to comment.
rucknium: My next step is to make a PR with the patch and the analysis I explained here. The full patch will remove dead logic, too.
vtnerd: the logic appears changed in this patch anyway https://github.com/monero-project/monero/commit/b750fb27b0f20e9443827732b69a504a76036430#diff-85f0577b849d28c11a899cb863aec3df1d946e0bb214a623deef05928271fef6
vtnerd: I believe the original logic was to attempt one whitelist, then grey each time
vtnerd: whereas this add the while loop
vtnerd: the original code may have been worse actually
https://libera.monerologs.net/monero-research-lab/20260930#c711528
Shi et al. (2026) "Are Unreachable Nodes Truly Safe? Fully Eclipsing Monero’s P2P Network!": https://arxiv.org/pdf/2609.10260
nevent1qqststnsjl4xtxtmuttamd34fg7umxrclwyqmzeutlsqm6j7y6t0rggrg663f