CVE-2026-4775: libtiff: integer overflow causes heap OOB write

A signed integer overflow in libtiff's putcontig8bitYCbCr44tile function lets a specially crafted TIFF file corrupt heap memory via incorrect pointer math, leading to a crash or potentially arbitrary code execution. This matters because libtiff sits quietly inside 129 downstream packages and, critically, inside Red Hat's AI Inference Server images (rhaiis/vllm-cuda-rhel9, vllm-rocm-rhel9, vllm-spyre-rhel9) — meaning any vision or multimodal inference stack built on those containers inherits the flaw whenever it decodes TIFF images. The actual urgency is moderate: EPSS sits at 0.55% (top 58th percentile), it's not in CISA KEV, there's no public exploit or Nuclei template, and CISA's SSVC decision is TRACK rather than Act. The attack vector is local with required user interaction, so exploitation needs a malicious file to be opened or auto-processed by an image pipeline rather than a bare network hit. Patch libtiff to the fixed version referenced in the RHSA errata (RHSA-2026:12265 and related), rebuild any custom rhaiis/vllm containers from patched base images, and if TIFF ingestion isn't required by your inference or RAG pipeline, disable it or restrict decoding to a sandboxed, resource-limited process.

No confirmed active exploitation of CVE-2026-4775 has been reported, but organizations should still patch proactively.

1) Patch libtiff to the version fixed in the referenced Red Hat errata (RHSA-2026:12265, 12271, 14929, 16055, 19150, 19363, 19585/86, 19604/08) across all affected packages, including compat-libtiff3 and mingw-libtiff builds. 2) Rebuild and redeploy any rhaiis/vllm-cuda-rhel9, vllm-rocm-rhel9, or vllm-spyre-rhel9 container images from the patched base rather than patching in place. 3) If your inference or RAG pipeline doesn't need TIFF support, strip or disable TIFF decoding entirely to eliminate the attack surface. 4) Where TIFF decoding is required, isolate it in a sandboxed/seccomp-restricted worker process with memory limits so a crash doesn't take down the whole inference service. 5) Detection: monitor image-decoding worker crash logs and core dumps for anomalous heap corruption signatures following TIFF uploads, and flag repeated malformed-TIFF submissions to the same endpoint as reconnaissance/exploitation attempts.

This vulnerability affects the following AI/ML architecture patterns: model serving, RAG pipelines, vision-language multimodal inference, data preprocessing pipelines.

CVE-2026-4775 has a CVSS v3.1 base score of 7.8 (HIGH). The EPSS exploitation probability is 0.55%.