# Eclipse Attacks on Ethereum's Peer-to-Peer Network

Crypto.world

> Eclipse attacks — where an adversary monopolizes all of a victim node's P2P connections — have been studied extensively for Bitcoin and Monero, but their practicality against Ethereum after the 2022…

**arXiv**: 2601.16560

**Published**: 2026-01-23

**Authors**: Ruisheng Shi, Yuxuan Liang, Zijun Guo, Qin Wang, Lina Lan, Chenfeng Wang, Zhuoyi Zheng

## AI summary

## Overview

**Research area:** Blockchain network security — specifically peer-to-peer (P2P) layer attacks on Ethereum execution-layer nodes in the post-Merge era.

**Technical level:** Advanced. The paper assumes familiarity with Kademlia-style DHTs, Ethereum's Discv4 discovery protocol, the ENR/node-record format, the devp2p crawler, and the split between execution and consensus clients.

**Scope (one sentence):** The paper designs, implements and evaluates a multi-stage eclipse attack that fully isolates a post-Merge Ethereum execution-layer node upon restart, using discovery-table poisoning, DNS peerlist infiltration, and network-wide connection-slot hijacking.

## What This Paper Is About

Eclipse attacks — where an adversary monopolizes all of a victim node's P2P connections — have been studied extensively for Bitcoin and Monero, but their practicality against Ethereum after the 2022 Merge was largely unexplored. The authors analyze how modern Geth nodes (specifically Geth v1.14.3, released May 9, 2024) discover peers, manage their discovery table and long-term database, and fill their 16 outgoing and 34 incoming connection slots. They then build an end-to-end attack that poisons those peer sources before a node restarts and floods its connection slots afterwards, and they test each stage on the Sepolia testnet and measure the mainnet.

## Key Contributions

1. **A systematic dissection of Ethereum's P2P stack** (Section 2): client architecture, the short-term discovery table versus the persistent DB, Discv4 message handling, the 17 K-buckets × 16 slots (272-node) structure, IP-based admission limits, connection lifecycle, rate limiting, and the devp2p crawler's scoring, update frequency and pruning behavior.
2. **The first end-to-end implementation of an eclipse attack against a post-Merge Ethereum execution-layer node**, structured as five steps: (i) DB pre-filling to poison the discovery table, (ii) DNS list poisoning, (iii) network-wide available-slot occupation, (iv) outgoing connection hijacking, and (v) incoming connection hijacking.
3. **A DNS peerlist poisoning technique that the authors describe as the first in the cryptocurrency context**, requiring only 28 IP addresses over 100 days and involving identification of the official devp2p crawler and manipulation of its scoring.
4. **Evaluation on both Sepolia testnet and Ethereum mainnet**, plus concrete countermeasures (Section 6) and responsible disclosure of all findings to Ethereum's security team.

## Main Findings

- **Discovery-table filling is fast and repeatable:** 272 attacking nodes raised attacker occupancy of the target's discovery table to 95% within 24 hours. In six repeat rounds (a new batch every 2 hours), each batch refilled the table to nearly 100% within about 1 hour. Table 1 reports per-trial fills ranging from 249/272 (91.54%) to 265/272 (97.43%).
- **DB growth is modest:** a freshly deployed Sepolia full node's DB grew at an average rate of 200 nodes per month; with Sepolia estimated at around 2,000 total nodes, the maximum realistic DB size on Sepolia is expected to reach 2,000 entries.
- **Poisoning concentrates in the last buckets:** seed nodes randomly selected from the DB are typically placed in the last 8 buckets, so later filling rounds target those 8 buckets with 128 (= 8 × 16) nodes each. Without any DB filling, the last 2 buckets are 0% filled and the last 5 buckets only 22% filled. As DB fill rose from 33% to 66%, last-two-bucket occupancy rose from 37% to 70% (unchanged attacker behavior) and from 36% to 68% (with behavior adaptation); last-five-bucket occupancy rose from 28% to 44% and from 22% to 76% respectively.
- **Bucket position determines hijack odds:** the reported outbound selection rates are 75% (last 2 buckets), 96.9% (last 5), 99.6% (last 8) and 100% (all 17).
- **The devp2p crawler can be located and fed:** the crawler IP was identified as 34.243.109.22, with a randomized key pair and UDP port at each startup. In a controlled testbed (no attack on the official crawler), 5 out of 10 experiments got attacker nodes into the target's discovery table, and in one experiment all five attacker nodes reached the maximum score of 3 points.
- **DNS list poisoning is slow but bounded:** mainnet nodes are likely included in the DNS list within 1–2 days, Sepolia nodes in about 10 days. With an estimated average daily score advantage of 5 points, Table 3 estimates 54 days to replace 0%, 64 days for 25%, 100 days for 50%, 166 days for 75%, and 538 days for 100% of entries (scores 275, 325, 506, 831 and 2,688 respectively).
- **Idle capacity is scarce:** across more than 2,000 probed public Ethereum nodes, over 80% limited available incoming slots to 10 or fewer.
- **Network-wide slot occupation works on Sepolia:** of 1,268 Sepolia nodes screened from the official DNS list on July 17, 2025, 321 had available slots; using 200 Sepolia nodes, about 90% of those slots were occupied after two hours, leaving 34 nodes still with availability.
- **Slot occupation amplifies outgoing hijacking:** with DB fill fixed at 50%, outgoing hijack success at 25% DNS poisoning rose from 4/20 (20%) to 6/20 (30%), and at 50% DNS poisoning from 8/20 (40%) to 19/20 (95%). The abstract summarizes the effect as raising outgoing redirection success from 45% to 95%.
- **Benign connections mostly come from DNS:** among 56 experiments where the target's incoming slots were not fully occupied, 13 (23%) had benign connections originating from the discovery table and 51 (91%) from the DNS list.
- **Incoming hijacking is near-certain on testnet, partial on mainnet:** Sepolia succeeded in 30/30 trials (100%), all within 30 seconds of restart. Mainnet succeeded in 24/40 (60%) within two days, with 18/40 (45%) within 30 seconds and 23/40 (57%) within one day.
- **Client update cadence widens the window:** between the Merge (Sep. 15, 2022) and May 5, 2025 — over 32 months — 56 Geth versions were released, averaging about 1.75 per month, each generally requiring a restart.
- **IP cost is bounded:** the authors estimate 208 public IPs for Sepolia DB pre-filling at 50% fill (2,000 nodes over 16 rounds of 128 nodes, 208 = 13 × 16) and 624 for mainnet (~6,000 nodes, i.e., 208 × 3); 28 IPs for DNS poisoning (272 / 10); 200 active nodes for slot occupation; 28 IPs for outgoing hijacking; and 40 IPs for incoming hijacking. Totals: 208 + 28 + 28 + 40 = **304** public IPs for Sepolia and 624 + 28 + 28 + 40 = **720** for mainnet.

## Methodology in Plain English

The authors first reverse-engineer how a Geth node finds and keeps peers, then attack each mechanism in turn.

- **Pre-poisoning the saved peer book.** A node keeps a persistent database of every node it has met. On startup it loads 30 random entries from that database plus 4 hardcoded bootnodes (34 seeds) and uses them to build a fresh in-memory discovery table. The attackers send unsolicited Ping messages so the victim adds them to the table; entries that survive 5 minutes and one liveness check get saved permanently. By cycling batches of attacking nodes — start a batch, let it get saved, kill it, start the next — the attackers stuff the database, then wait for a second restart so the victim seeds itself mostly with attacker entries.
- **Poisoning the official DNS peer list.** Ethereum ships a list of peer records maintained by a crawler. The authors find the crawler, make their nodes easy to rank highly (low node IDs prefixed with zeros, so they sort first in the crawler's input list), infer the crawler's schedule by watching for ENRRequest messages, and answer its requests to accumulate score points until they enter the top-N list (3,000 for mainnet, 250 for testnet). They did not attack the real crawler; they replicated it on two servers.
- **Occupying everyone else's slots.** Because each node accepts only 34 incoming connections and no active eviction exists, the attackers pre-fill honest nodes' free incoming slots. When the victim restarts, honest peers are already full and cannot accept the victim's outgoing connection attempts.
- **Hijacking the victim's own connections.** On restart the victim's table is refilled by attacker Pings and DNS entries, so its 16 outgoing connections go to attackers; simultaneously, many attacker nodes open inbound connections to fill its 34 incoming slots.
- **Measurement.** Each stage was run as a controlled experiment (Sepolia, mostly, with mainnet measurements for slot availability and incoming hijacking), with the target's IP-based admission and 30-second per-IP rate limits disabled where noted to conserve IP resources.

## Why This Matters

For research, the paper moves eclipse attacks on Ethereum from pre-Merge theory into a demonstrated, costed exploit against the current stack, and shows that post-Merge design choices — long-lived validators, static bootstrapping, heavy DNS reliance, dominance of a few client pairings such as Prysm+Geth — make the attack *easier* rather than harder. It also introduces DNS peerlist poisoning as a general technique that other cryptocurrency networks relying on crawled peer lists should evaluate.

Real-world implications:

- **Validator operators** face a recurring restart window (56 Geth releases in 32 months, roughly 1.75 per month) during which their node can be captured by an adversary and fed false mempool or transaction data.
- **Execution-client users and RPC providers** depend on honest peer sets for transaction gossip; isolation breaks that assumption without touching consensus finality.
- **Testnet and small-network operators** are the most exposed: on Sepolia the attack succeeded in 30/30 incoming-hijack trials within 30 seconds, versus 24/40 on mainnet.
- **Client developers** gain a concrete checklist — ping-rate blacklisting, DNS IP-per-subnet caps, DNS node reporting/blacklists — that maps directly onto code they maintain.

Industry relevance is high for staking services, node-hosting providers and client teams, because the entire attack was disclosed to Ethereum's security team and the mitigations are implementation-level rather than requiring a consensus change.

## Future Directions

- **Measuring real-world crawler manipulation at scale.** The DNS experiment was a two-server replica; the open question is how much faster (or slower) a live campaign against the official crawler would be than the 100-day, 28-IP estimate.
- **Evaluating the proposed countermeasures.** The paper proposes blacklisting peers that send more than 5 Pings within 1 minute (initial 1-minute ban, doubling to 2, 4, 8 minutes up to a cap such as 24 hours) and limiting or blacklisting DNS entries by IP, but the truncated content does not report an evaluation of their effectiveness or their cost to honest nodes.
- **Extending the analysis to consensus-layer clients and other client pairings.** The paper scopes itself exclusively to the execution client, leaving the beacon-chain P2P stack and the dominance of specific pairings like Prysm+Geth as open ground.
- **Understanding the 80%-low-idle-capacity finding.** Over 80% of public nodes keep 10 or fewer free incoming slots; whether raising default capacity, adding active eviction, or prioritizing inbound connections from diverse subnets would meaningfully reduce eclipse feasibility is not resolved here.

## Target Audience

Blockchain security researchers and graduate students working on P2P-layer attacks; Ethereum client developers and protocol designers evaluating peer-management hardening; staking operators and node-hosting providers who need to understand restart-time exposure; and network-measurement practitioners interested in crawler-based peer discovery systems. Readers without background in Kademlia-style DHTs or Ethereum's Discv4/devp2p stack will find Section 2 a necessary prerequisite.

## Authors’ abstract

Eclipse attacks isolate blockchain nodes by monopolizing their peer-to-peer connections. The attacks were extensively studied in Bitcoin (SP'15, SP'20, CCS'21, SP'23) and Monero (NDSS'25), but their practicality against Ethereum nodes remains underexplored, particularly in the post-Merge settings. We present the first end-to-end implementation of an eclipse attack targeting Ethereum (2.0 version) execution-layer nodes. Our attack exploits the bootstrapping and peer…

[The full abstract on arXiv](https://arxiv.org/abs/2601.16560)

[Read the original paper](https://arxiv.org/abs/2601.16560)

---

Canonical: https://crypto.world/paper/eclipse-attacks-on-ethereum-s-peer-to-peer-network-2601.16560
