smg multimodal CVE-2026-116095

The application must expose a code path that passes attacker-supplied JPEG data into the llm-multimodal component's jpeg_turbo::decode_jpeg_rgb function (e.g. an image-ingestion or multimodal inference endpoint accepting uploaded images). Per the CVSS vector exploitation is remote, unauthenticated (PR:N), and needs no user interaction (UI:N) against default handling of such images. What LIMITS impact: the consequence is availability-only (VA:L / A:L) memory exhaustion - there is no code execution, data disclosure, or integrity impact - and the pure-Rust fallback decoder already enforced a 512 MiB cap, so only the libjpeg-turbo decode path (where dimensions up to 65500×65500, ~12 GiB, are accepted) is affected by the unbounded allocation. The attacker needs a reachable path that actually decodes the malicious image for the bug to trigger.

This is a medium-severity denial-of-service issue, not a high-priority emergency. The vendor CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N) scores 6.9 and shows the flaw is remotely reachable with low complexity and no authentication or user interaction, but the impact metrics are narrow: VC:N and VI:N (no confidentiality or integrity impact) with only VA:L (low availability impact) - consistent with CWE-789 uncontrolled memory allocation causing resource exhaustion rather than code execution or data leakage. The exploit-maturity signal E:P (proof-of-concept) aligns with the input, which shows a public PoC reference (a GitHub gist) that is NOT verified as working exploit code. There is no CISA KEV listing and no EPSS score was provided, so there is no evidence of active or widespread exploitation; the vuln.today decision label is 'this_month', indicating routine-priority patching rather than act_now/emergency. The one tension worth flagging: the input Tags include 'Information Disclosure', which directly contradicts the CVSS (VC:N) and the CWE (uncontrolled allocation) - this is a DoS/availability bug, not disclosure. Net: a legitimately remote-reachable DoS in a specific image-decoding path, but low real-world urgency given availability-only impact, an unverified PoC, no KEV, and an obscure product footprint.

An attacker submits a crafted JPEG whose header declares enormous pixel dimensions to any feature of the application that feeds images into the llm-multimodal decoder (jpeg_turbo::decode_jpeg_rgb). Because the code computed the RGB buffer size (width × height × 3) with only overflow checking and no upper bound, the decoder attempts a multi-gigabyte allocation, exhausting memory and denying service. A public PoC reference (gist) exists but is not verified as working exploit code; exploitation is remote and requires no authentication per the CVSS vector.