FFmpeg CVE-2026-65705

Exploitation requires the target to run the FFmpeg vf_floodfill filter with filtergraph reinitialization explicitly disabled via the -reinit_filter 0 option - a non-default configuration - and to process a dynamically sized video stream in which a frame larger than the initial frame arrives after config_input() has fixed the traversal stack size. The attacker must get the victim to process attacker-supplied media (CVSS UI:P passive interaction), and the attack is local (AV:L), not remote/network-triggered. Limiting factors: the -reinit_filter 0 setting is not the default (normal reinitialization prevents the overflow), the floodfill filter must be present in the filtergraph, and the input must actually change dimensions upward; PR:N means no authentication is needed once the victim processes the file. These constraints substantially narrow real-world exposure below what the raw score alone implies.

The provided CVSS 4.0 vector (AV:L/AC:L/AT:P/PR:N/UI:P/VC:H/VI:H/VA:H) scores 7.3 and tells a consistent story of meaningful-but-constrained risk: exploitation is local (the attacker supplies a file rather than reaching the process over a network), requires no privileges, but depends on both a passive user interaction (processing the media) and an attack requirement (AT:P) - here the non-default -reinit_filter 0 setting combined with a size-changing stream. There is no CISA KEV listing and no public exploit was identified at time of analysis, so this is not confirmed actively exploited; the high VC/VI/VA impact reflects potential code execution, but the description confirms only reliable heap corruption and crash, with true RCE contingent on heap layout and hardening. Net: a real memory-safety bug worth patching, but the non-default configuration requirement and local/UI vector make it a lower operational priority than a remote unauthenticated RCE with the same CVSS. No EPSS score or SSVC decision data was provided, which limits automated prioritization confidence.

An attacker crafts a video file whose stream begins with small frame dimensions and later switches to larger dimensions, then delivers it to a victim or service that transcodes it through an FFmpeg filtergraph using the floodfill filter with -reinit_filter 0. When the larger frame reaches filter_frame(), neighbor pushes overflow the undersized traversal stack, corrupting the heap and crashing the process, with potential for code execution depending on heap layout. No public exploit was identified at time of analysis, and the local vector with passive user interaction means the attacker relies on the target choosing to process the malicious media.