CVE-2026-43628

This vulnerability in the DRY sampler's boundary-checking logic is more dangerous than its CVSS 7.8 suggests. When dry_allowed_length is set to INT32_MIN, the comparison intended to prevent out-of-bounds access wraps around and becomes the vector for reading far beyond the allocated buffer. The result is either a SIGSEGV crash (obvious, self-correcting) or probability corruption (silent, compounding). The probability corruption vector is what you should prioritize. An attacker who calibrates their payload to avoid crashing can inject garbage reads into token sampling without triggering any alerts. For multi-tenant inference deployments, this creates a lateral contamination risk: one user's crafted request can subtly degrade inference quality for subsequent users sharing the same context window and cached KV state. This isn't a crash you notice — it's silent inference integrity compromise. Affected builds span b3978 through b9058. Check your deployed version. The DRY sampler is an internal constraint mechanism that was apparently exposed as an HTTP API parameter — this architectural decision is the root exposure, not merely the integer underflow. Your HTTP API layer likely never validated this parameter against the semantic constraints of the sampler, treating it as an opaque passthrough. Immediate actions: (1) Verify your deployed version is outside the affected range or patched; (2) Audit your HTTP API layer for sampler parameters that are exposed without semantic validation — if dry_allowed_length was exposed, other sampler configuration likely was too; (3) If you cannot patch immediately, consider-rate limiting or rejecting requests where sampling parameters approach integer boundary values; (4) Monitor for subtle quality degradation in inference outputs that doesn't correspond to input changes, as this may indicate the corruption path is active.

Reviewed through automated stages and approved by a human before publication.