GHSA-MX64-MJ3Q-7PRJ

Memory Exhaustion via Unbounded Map Allocations in Avro Decoder

The Avro map decoder accepted attacker-controlled block-element counts from the wire format and grew the destination map without enforcing an upper bound. The slice decoder already had Config.MaxSliceAllocSize for the equivalent attack against arrays; the map decoder had no analogous limit, so a producer could declare an arbitrarily large map (in one block, or chunked across many sub-limit blocks) and exhaust process memory until the OOM killer fired.

The fix introduces Config.MaxMapAllocSize with cumulative enforcement across block boundaries. The new limit is opt-in: the field defaults to zero, which preserves the previous unbounded behavior for backward compatibility. Upgrading to v2.33.0 alone does not mitigate the issue, consumers of untrusted Avro data must explicitly set MaxMapAllocSize on their avro.Config.

Description

Avro maps are encoded as a sequence of blocks; each block declares a long element count followed by that many key/value pairs. The decoder uses these counts both to size the destination map and as the loop bound for reading entries.

Pre-fix, the map decoder enforced no upper limit at any layer:

• No per-block element-count check.
• No cumulative across-block element-count check.
• No memory-budget check before make(map[...]..., n) or before growing the map.

The slice decoder had been hardened via Config.MaxSliceAllocSize and tracked cumulatively across blocks; the map decoder was a missing-by-symmetry gap. Even a partial per-block bound on maps would have been insufficient on its own, Avro permits encoding a logical map as many small blocks, so a producer could split a 10 GB map into 10,000 sub-MaxMapAllocSize blocks and still drive total allocation past any single-block threshold. The fix tracks cumulative entry count at block-header boundaries, before the block's entries are decoded into the map, and errors out before allocation when the running total would exceed the configured cap.

Two decoder variants were affected, both in codec_map.go:

• mapDecoder.Decode, string-keyed maps.
• mapDecoderUnmarshaler.Decode, encoding.TextUnmarshaler-keyed maps (e.g. map[CustomKey]V where *CustomKey implements UnmarshalText).

Affected components

PR #5 (fix/map-alloc-chunking-bypass) covers both decoders and adds chunking-attack tests for both. The same PR also adds the previously-missing chunking-attack test coverage for the slice path in 534c7518, the slice logic was already correct, only its test coverage was incomplete.

Technical details

The fix mirrors the slice decoder's pattern:

1. At each block header, read the element count as int64.
2. Add it to a running total maintained across the block loop.
3. If the running total exceeds Config.MaxMapAllocSize (when nonzero), return an error before allocating any of that block's entries.
4. Otherwise, decode the block's entries into the map.

Per-block enforcement alone would be bypassable by chunking; cumulative tracking closes that. The check sits at the block-header read, before per-entry allocation, so a single oversized block also cannot allocate first and then fail post-hoc.

Config.MaxMapAllocSize semantics match Config.MaxSliceAllocSize: zero means unlimited, any positive value is the cumulative cap on element count (not byte size).

Fixed behavior

v2.33.0 adds the MaxMapAllocSize configuration field and the cumulative-enforcement logic in both map decoders. Both decoders return a descriptive error when the cumulative entry count would exceed the configured cap; no entries are allocated past the limit.

Tests added in PR #5 cover, for both mapDecoder and mapDecoderUnmarshaler:

• Single-block allocation exceeding the limit (rejected before allocation).
• Chunking attack: multiple sub-limit blocks whose cumulative count exceeds the limit (rejected at the block-header that crosses the threshold).
• Multi-block under the limit (decoded normally).

Affected versions

• github.com/hamba/avro/v2, all versions up to and including v2.31.0 (repository is read-only upstream).
• github.com/iskorotkov/avro/v2, all versions prior to v2.33.0. Note: v2.33.0 and later are vulnerable by default and only protected when MaxMapAllocSize is explicitly configured, see Mitigation.

Fixed versions

github.com/iskorotkov/avro/v2 v2.33.0 and later, with Config.MaxMapAllocSize explicitly set to a non-zero value.

A bare upgrade to v2.33.0 without setting MaxMapAllocSize leaves the decoder in the same unbounded state as v2.32.0. This is a backward-compatibility choice; a future major version may flip the default. Until then, treat this advisory as requiring both an upgrade and a configuration change.

There is no upstream fix for github.com/hamba/avro/v2, module path is archived. Migrate to the fork as described under Mitigation.

Mitigation

Migrate from github.com/hamba/avro/v2 to github.com/iskorotkov/avro/v2 >= v2.33.0 and configure an allocation cap appropriate for your schema. The recommended approach for processes that decode untrusted input is a dedicated frozen config, used at every relevant call site, rather than mutating avro.DefaultConfig:

cfg := avro.Config{
    MaxByteSliceSize:  102_400,
    MaxSliceAllocSize: 10_000,
    MaxMapAllocSize:   10_000,
}.Freeze()

decoder := cfg.NewDecoder(schema, reader)

Choose the values based on the largest legitimate map your schema produces; a value 2–10× that ceiling provides headroom for benign variance while still bounding worst-case memory.

For consumers that prefer the original import path, a replace directive in go.mod is supported:

replace github.com/hamba/avro/v2 => github.com/iskorotkov/avro/v2 v2.33.0

replace is honoured only for the main module of a build, transitive consumers must add their own replace, or migrate the import path directly.

If you cannot upgrade immediately, the only structural workarounds are out-of-band: run decoders in memory-constrained child processes or cgroups so an OOM is contained, reject inputs from sources without resource controls, and apply per-request decode deadlines so a runaway decode at least times out before the OOM killer fires.

Proof-of-concept input

Two attack shapes, both targeting map[string]int:

Single-block, oversize block count. Emit one block header declaring n = 2³¹ − 1 (or any value whose n × averageEntrySize exceeds available memory) followed by truncated entries. Pre-fix, the decoder pre-allocates make(map[string]int, n), which fails or stalls long before EOF is reached.

Chunking bypass. Emit k blocks each declaring n / k elements, with n / k below any plausible per-block threshold but n itself well into the GB range. Pre-fix, the decoder happily grows the map block-by-block until the OS kills the process. Post-fix with MaxMapAllocSize = 10_000, the decoder rejects whichever block-header read pushes cumulative count past 10,000.

Either shape can be produced by hand-crafting the wire bytes; no iskorotkov/avro writer is needed to generate them.

References

• Fix PR: iskorotkov/avro#5
• Fix commit: 5192df9 (codec_map.go, config.go, tests)
• Slice-path chunking-attack test coverage added in the same PR: 534c7518
• Release: v2.33.0
• Security policy: SECURITY.md
• Related advisories on this fork: GHSA-mc57-h6j3-3hmv (integer overflow), GHSA-w8j3-pq8g-8m7w (CPU exhaustion, the same chunked-payload shape may trigger both before allocation pressure kicks in)
• Cross-module precedent on hamba/avro: GO-2023-1930 / CVE-2023-37475 / GHSA-9x44-9pgq-cf45
• Upstream (read-only): hamba/avro

Credits

• Fix author (commit 5192df9, PR #5, MaxMapAllocSize config field, cumulative enforcement in both map decoders, chunking-attack tests for slices and maps): Ivan Korotkov (@iskorotkov)
• Review (commit a5fbddcb, "address review comments"): Daniel Błażewicz (@klajok)

Timeline

• 2026-04-30, MaxMapAllocSize introduced (5192df9); chunking-attack test coverage for slices added (534c7518).
• 2026-05-01, PR #5 merged into main.
• 2026-05-06, v2.33.0 tagged and released.
• 2026-05-07, Advisory published.
• 2026-05-15, Advisory revised.

Crafted input forces the application to consume excessive CPU, memory, or other resources, degrading or denying service. Typical impact: denial of service.

GHSA-MX64-MJ3Q-7PRJ has a CVSS score of 7.5 (High). The vector is network-reachable, no privileges required, and no user interaction. A CVSS score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether this affects your application depends on whether the vulnerable code is present and reachable in your environment. A fixed version is available (2.33.0); upgrading removes the vulnerable code path.

Kodem intelligence

1. What is GHSA-MX64-MJ3Q-7PRJ? GHSA-MX64-MJ3Q-7PRJ is a high-severity uncontrolled resource consumption vulnerability in github.com/iskorotkov/avro/v2 (go), affecting versions < 2.33.0. It is fixed in 2.33.0. Crafted input forces the application to consume excessive CPU, memory, or other resources, degrading or denying service.
2. How severe is GHSA-MX64-MJ3Q-7PRJ? GHSA-MX64-MJ3Q-7PRJ has a CVSS score of 7.5 (High). This score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether it represents real risk in your environment depends on whether the vulnerable code is present and reachable.
3. Which versions of github.com/iskorotkov/avro/v2 are affected by GHSA-MX64-MJ3Q-7PRJ? github.com/iskorotkov/avro/v2 (go) versions < 2.33.0 is affected.
4. Is there a fix for GHSA-MX64-MJ3Q-7PRJ? Yes. GHSA-MX64-MJ3Q-7PRJ is fixed in 2.33.0. Upgrade to this version or later.
5. Is GHSA-MX64-MJ3Q-7PRJ exploitable, and should I be worried? Whether GHSA-MX64-MJ3Q-7PRJ is exploitable in your environment depends on whether the vulnerable code is present and reachable. A CVSS score is a worst-case rating; it does not account for your specific deployment, configuration, or usage patterns. Kodem, an Intelligent Application Security platform, uses runtime intelligence to show which vulnerabilities actually execute in production, so you can focus on the ones that represent real risk. Get a demo
6. What actually determines whether GHSA-MX64-MJ3Q-7PRJ is exploitable, and how bad it is? Exploitability and impact are not fixed properties of a CVE. They depend on runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A high CVSS score on a dependency that never runs is not the same as real risk. Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter.
7. How do I fix GHSA-MX64-MJ3Q-7PRJ? Upgrade github.com/iskorotkov/avro/v2 to 2.33.0 or later.