
RedisBloom CMSketch RDB-load integer overflow (CVE-2026-62356) — non-weaponized PoC: deterministic crash/DoS reproducer, analysis & version matrix
An integer-overflow vulnerability in RedisBloom's Count-Min Sketch (CMSketch) RDB
deserialization path. An authenticated Redis client can load a specially crafted
CMSketch object (via RESTORE or a crafted RDB file) whose declared dimensions
overflow the expected-buffer-length computation, causing an undersized heap
buffer to be accepted. Any subsequent CMS command (CMS.QUERY / CMS.INCRBY /
CMS.INFO) then indexes far out of bounds — a crash (DoS) at minimum, and, per
the upstream advisory, potentially remote code execution in the context of the
Redis server process.
This repository contains a non-weaponized proof of concept: a deterministic crash/DoS reproducer plus supporting analysis tooling. It does not contain a working RCE exploit chain.
| CVE | CVE-2026-62356 |
| GHSA | GHSA-q5p6-hwxh-c23h |
| Reporter (to upstream) | josephsurin |
| CWE | CWE-190 (Integer Overflow), CWE-787 (OOB Write) |
| Published | 2026-08-17 |
| CVSS | High (CVSS:4.0 AV:A/AC:H/PR:L/…) |
Affected versions (per upstream advisory): Redis OSS 8.0.0–8.0.8, 8.2.0–8.2.15, 8.4.0–8.4.8, 8.6.0–8.6.5, 8.8.0–8.8.1, 8.10.0; RedisBloom 2.0.0–2.8.23. Fixed in: Redis 8.8.2 / RedisBloom 2.8.24 (and corresponding backports).
This repository is an independent reproduction. Credit for discovery and responsible disclosure belongs to the upstream reporter named above.
CMSRdbLoad reads attacker-controlled width / depth from the serialized
object and computes the expected array length as 4 * width * depth in 64-bit
arithmetic without an overflow check. Because the payload length check
compares against this wrapped product, a payload whose dimensions make the
product wrap (e.g. width = depth = 2^32 → product = 2^66 mod 2^64 = 0) passes
validation with an empty or undersized array buffer, while the in-memory
sketch retains the huge declared dimensions.
Every later cell access is:
loc = (MurmurHash2(item, seed = row) % width) + row * width;
array[loc] ... // array is the undersized buffer; loc is huge
so any CMS command dereferences array + 4*loc — an out-of-bounds access at an
attacker-influenceable offset (the hash is computable offline, MurmurHash2 with
known seeds). The full analysis is in docs/ANALYSIS.md.
Environment: Docker, any host platform (use --platform linux/amd64 on ARM Macs).
# 1) start a vulnerable lab (redis:8.8.1, bundled pre-fix RedisBloom, port 6381)
bash poc/setup_lab.sh
# 2) inject the poisoned CMSketch object and trigger the crash
python3 poc/cms_overflow_crash.py --port 6381
# -> CMS.QUERY gets no reply; subsequent PING fails: SIGSEGV confirmed
# -> server log shows "Accessing address: 0x..." inside CMS_Query/CMS_IncrBy
# 3) (optional) verify persistence: poisoned key survives restart via RDB
docker restart cve-2026-62356-lab
# any CMS.QUERY on the key crashes the server again — persistent DoS
# 4) cleanup / recover
python3 poc/cms_overflow_crash.py --port 6381 --cleanup
Expected output (step 2):
[*] sanitize-dump-payload: +OK
[*] injecting poisoned CMS key (width=depth=2^32, array=NULL)...
[*] RESTORE -> +OK
[*] predicted fault address (depth=1): 0x...
[*] triggering via CMS.QUERY ...
[*] CMS.QUERY -> (no reply / crashed)
[*] PING -> (no reply / crashed)
poc/
setup_lab.sh start a pinned vulnerable redis:8.8.1 container
cms_overflow_crash.py crash/DoS reproducer (inject + trigger + cleanup)
cmslib.py shared RESP client / payload builder helpers
leak_base.py analysis tool: crash-free heap-layout inference (read-only)
mkwin2.c analysis tool: MurmurHash2 preimage table generator (offline)
mm2_preimage.c analysis tool: single-hash MurmurHash2 preimage solver (offline)
docs/
ANALYSIS.md detailed vulnerability analysis (English)
RESTORE, MIGRATE, DEBUG and RDB-file
ingestion to trusted users via ACLs; treat any serialized-data import path
(replication, --rdb preload, backup restore) as equally sensitive.This repository is published for educational and defensive research purposes only, to help operators understand, detect, and patch the vulnerability.
Discovery and disclosure credit: josephsurin (upstream reporter). This repository's reproduction and analysis are independent work.