FFmpeg CVE-2026-70630

Exploitation requires FFmpeg to process a crafted file that contains an SPV1 (Screenpresso) encoded video track - specifically one where the zlib payload decompresses to fewer bytes than the full frame pixel buffer. The CVSS 4.0 vector specifies UI:P, meaning a user or automated process must actively trigger decoding of the malicious file; passive exposure is not possible. AV:L indicates the attack is scoped to the process's local context, though in practice any service that accepts untrusted media uploads and decodes them with FFmpeg effectively exposes this attack surface over the network. No authentication is required to supply the crafted file if the decoding service is publicly accessible. Exploitation is limited to: (1) environments where FFmpeg is compiled with Screenpresso decoder support (enabled by default in standard builds), (2) FFmpeg versions 3.0-8.x, and (3) scenarios where the attacker can retrieve or observe the decoded output (e.g., a thumbnail returned by the service). Environments using FFmpeg only as a command-line tool by local users with no untrusted-input pipelines face minimal realistic risk.

The CVSS 4.0 base score of 6.8 (AV:L/AC:L/AT:N/PR:N/UI:P/VC:H) accurately captures the core risk profile: local attack vector with required user interaction (the target must process the crafted file), no privileges needed, and high confidentiality impact in the vulnerable system with no integrity or availability dimension. The local attack vector is the most significant limiting factor - exploitation requires the attacker to cause FFmpeg to process a malicious file, whether by convincing a user to open it, injecting it into an automated transcoding pipeline, or delivering it via a media upload endpoint. For batch-processing and thumbnail-generation services this effectively becomes network-triggerable with only AV:L at the FFmpeg process level. KEV is absent and no POC has been identified, making this currently a patch-priority rather than emergency-response item; however, the VC:H rating is credible given that heap leak primitives can facilitate ASLR bypass in memory-safe exploit chains. EPSS data was not provided in the input and cannot be stated. No conflicting signals between CVSS vector and description are present.

An attacker crafts a Screenpresso (SPV1) video file whose zlib-compressed payload decompresses to fewer bytes than the declared frame dimensions require - for example, a 1920×1080 frame whose payload expands to only half the expected byte count. When submitted to a vulnerable FFmpeg-based transcoding service, thumbnail extractor, or media player, screenpresso_decode_frame() inflates the short payload and then copies the entire full-frame-sized buffer into the output picture, including heap residue from prior allocations. In a persistent service that has previously processed other media or handled authentication tokens, this residue may contain userspace pointers or credential material, which appear embedded in the decoded video frames returned to the attacker.