Apache DataSketches CVE-2026-103634

Exploitation requires that the target application calls count_min_sketch::deserialize() on serialized Count-Min sketch data originating from an untrusted or attacker-influenced source; the vendor explicitly states 'Only applications that deserialize Count-Min sketches from untrusted sources are affected.' The affected code path is datasketches-cpp versions 4.1.0 through 5.2.0. The over-read is triggered by a truncated sketch (missing preamble bytes), while the heap OOB read/write requires a crafted sketch whose serialized bucket count and hash-function count overflow 32-bit arithmetic to zero; the resulting corruption manifests only on subsequent update or estimate-query operations, not at deserialization time. The principal limiting factor is that deserializing sketches from untrusted input is not a default or universal usage pattern - closed pipelines that only process internally generated sketches are not exposed; versions 5.3.0 and later are not affected.

This is a memory-corruption flaw (out-of-bounds read and write) in Count-Min sketch deserialization in Apache DataSketches C++ (datasketches-cpp), rated moderate by the vendor. No CVSS vector, EPSS score, or CWE was supplied in any source, so quantitative exploitation-probability signals are unavailable. There is no CISA KEV listing and no public exploit code referenced, so exploitation status is 'no public exploit identified at time of analysis.' The root causes are concrete and well-described: deserialize() omits the preamble from its input-size check (up to a 16-byte over-read), and a 32-bit integer overflow in the bucket/hash-function table-size computation can wrap to zero, producing an empty table whose later updates and estimate queries read and write outside the heap allocation. Real-world risk is strongly gated by one precondition: the application must deserialize Count-Min sketches from untrusted sources - not a default behavior of most deployments. My independent CVSS 3.1 assessment (base ~7.3, High) comes out higher than the vendor's 'moderate' rating; the likely reason for the gap is that the vendor weighted the untrusted-deserialization precondition and the fact that reliable arbitrary code execution is only described as 'potentially' achievable, not demonstrated. Treat this as a genuine but conditional priority: urgent for services that ingest serialized sketches from external/untrusted parties, low-priority for closed pipelines that only process self-generated sketches. Authentication requirements cannot be confirmed because no CVSS vector was provided.

An attacker supplies a crafted serialized Count-Min sketch to a service or pipeline that calls count_min_sketch::deserialize() on externally sourced data. A truncated preamble triggers a read up to 16 bytes past the input buffer, or manipulated bucket/hash-function counts overflow 32-bit arithmetic to zero, yielding an empty table whose subsequent updates and estimate queries corrupt heap memory, causing a crash and potentially enabling further exploitation. No public exploit code has been identified at time of analysis.