DataSketches C++ CVE-2026-103501

Exploitation requires that the target application invoke Apache DataSketches C++ HLL sketch deserialization (from a byte buffer or stream) on data from an untrusted source, and specifically that an attacker-controlled sketch be in LIST mode with a coupon count larger than the fixed 8-entry buffer. The key limiting factor is application design: the vulnerable code is a library routine with NO inherent network listener, so deployments that only deserialize internally generated or trusted sketches are not realistically exposed despite the AV:N vector. No authentication or user interaction is required at the library level (PR:N/UI:N), and the default code path performs no capacity check, so no non-default configuration is needed to trigger the bug once untrusted input reaches the parser. Affected versions 1.0.0-incubating through 5.2.0; fixed in 5.3.0.

Signals conflict and must be weighed carefully. NVD assigns CVSS 9.8 (AV:N/AC:L/PR:N/UI:N, C:H/I:H/A:H), but Apache's own release notes rate this 'moderate', and the description frames it as a library-level deserialization flaw affecting ONLY applications that deserialize HLL sketches from untrusted sources - so the AV:N/PR:N vector is only realized when a consuming application actually exposes sketch deserialization to attacker-controlled input. EPSS is low at 0.21% (11th percentile), there is no CISA KEV listing and no public exploit identified, consistent with a newly disclosed library bug rather than active or widespread exploitation. The demonstrable impact is a controlled heap write of up to 988 bytes (CWE-122) causing memory corruption and crash; the description states code execution is only 'potentially' enabled, so the full C:H/I:H claim is unproven rather than confirmed. Net: this is a genuine high-severity memory-corruption bug worth patching promptly wherever the library processes untrusted input, but the headline 9.8 overstates universal real-world risk - exposure is gated on application design, exploitation beyond DoS is unproven, and the vendor's 'moderate' rating plus low EPSS argue against treating it as an emergency for deployments that only process trusted, internally generated sketches.

An application (e.g. a data-analytics service) accepts serialized Apache DataSketches HLL sketch images from untrusted clients and deserializes them. An attacker submits a crafted LIST-mode sketch whose coupon count far exceeds the 8-entry internal buffer, writing up to 988 bytes past the heap buffer, corrupting heap memory and crashing the process - with potential for further exploitation toward code execution. No public exploit code is identified at time of analysis, and the AV:N/AC:L vector makes triggering straightforward wherever untrusted sketch input is reachable.