CVE-2026-70629

The CVSS 5.5 rating misrepresents this vulnerability by treating it as an isolated memory read rather than a systematic probing mechanism against persistent decoding state. This is worth your attention if you run any long-running FFmpeg-based media service. The technical root cause is a persistent buffer (ctx->inflated_buf) in the RSCC decoder that gets reused across decode operations instead of being reinitialized. When an attacker sends a crafted RSCC stream with mismatched tile geometry, the decoder copies uninitialized heap contents into the output frame. That's the trigger. But the vulnerability isn't a one-shot—each malformed input can probe a different slice of heap memory, and because the buffer persists across frames, the attacker can accumulate different heap state over time. This is heap archaeology, not a single-frame disclosure. The realistic threat model is a multi-tenant media service: a video transcoding daemon processing client uploads, a surveillance system decoding sequential streams, any service that reuses the same codec context across unrelated jobs. What the attacker reads depends entirely on what the service was doing moments before their input arrived. If your transcoding pipeline processes Client A's video, then Client B's, then an attacker submits a malformed RSCC stream—the read exposes whatever residual data from Client A's processing happened to occupy that heap region. Two things you should verify: First, check whether your service reinitializes FFmpeg codec contexts between independent processing jobs, or whether you maintain persistent contexts for performance. Second, determine if your demuxer pipeline validates tile geometry parameters before passing them to the RSCC decoder, or whether malformed parameters pass through unfiltered. If you can inject the tile geometry trigger through a standard container format (MP4, MOV), the attack is straightforward. If it requires raw codec bitstream injection, the complexity rises. The downstream risk extends beyond the initial read. If your pipeline writes the decoded frame to any client-accessible location—thumbnails, transcoded output, analytics logs—the contaminated frame enters every subsequent processing stage. A heap disclosure becomes a data exfiltration path through your legitimate output channels.

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