AISLE Discovers 6 High and Critical CVEs in FFmpeg

AISLE has discovered six additional high- and critical-severity vulnerabilities in FFmpeg, the engine that powers online video processing systems like YouTube and Netflix. Coming on the heels of the 21 security issues and 6 CVEs we discovered in late July, these findings are the second batch AISLE has surfaced in a codebase that both Google and Anthropic scanned earlier this year.

FFmpeg is part of the code the modern web runs on, and it’s been hardened over the course of decades. That makes it a true challenge for AI code analyzers: the “low-hanging fruit” was picked long ago.

In total, AISLE discovered 7 CVEs in August, with 6 of them receiving high or critical CVSS severity scores. Among them was a stack buffer overflow that had persisted in FFmpeg code for 19 years, surviving repeated rounds of fuzzing, static analysis, and expert human review.

This shows that AISLE's multi-model engine does much more than find simple bugs at scale: it delivers the sort of detailed analysis that experienced specialists struggle with.

6 High or Critical Vulnerabilities

Far from being low-hanging fruit, all of these CVEs were found in less-traveled protocol handlers: RIST, MPEG-PS, HEVC muxing, VC-2/Dirac and AV1 RTP packetizing, and DASH. Moreover, three of them had been hiding in FFmpeg code for nine or more years.

And as we’ve found on our road to finding 350+ CVEs with AI in 12 months, that is to be expected. AISLE’s AI code analyzer excels at inspecting the places that busy human reviewers tend to overlook, allowing vulnerabilities to persist in even rigorously inspected codebases.

Here’s what AISLE found:

CVE-2026-75143 (9.8 Critical) — Heap buffer overflow in the RIST protocol reader (libavformat/librist.c). librist_read() ignored the caller's buffer-size argument and copied the entire received payload, overflowing the destination buffer. Remotely triggerable via the async:rist:// URL scheme with no auth or user interaction. Fixed in commit 1c10bcc2.

CVE-2026-75141 (7.8 High) — Heap buffer overflow in the hvcC box writer when muxing HEVC: writing more NAL units of one type than the count field can hold overflows the counter and the heap buffer. Fixed in commit acf5d7cd.

CVE-2026-75142 (7.8 High) — Stack buffer overflow in the MPEG-PS muxer (libavformat/mpegenc.c) when input has more streams than the fixed-size buffer can hold. Introduced in 2007, this 19-year-old bug was fixed in commit 9d786e4b.

CVE-2026-75144 (7.8 High) — Heap buffer overflow in the VC-2/Dirac RTP packetizer from missing bounds checking on crafted input. Fixed in commit 1cdeb3c4.

CVE-2026-75146 (8.1 High) — Out-of-bounds read in the DASH demuxer (libavformat/dashdec.c): a live manifest refresh with a lower startNumber causes negative array indexing. Fixed in commit 65b0dab9.

CVE-2026-75147 (7.1 High) — Out-of-bounds read in the AV1 RTP packetizer's keyframe-detection loop: an unbounded OBU size causes pointer wraparound, dereferencing memory past the packet buffer. Fixed in commit 983dae9c.

One medium-severity finding

CVE-2026-75145 (5.8 Medium) — Incorrect integer narrowing conversion in the AV1 RTP packetizer: an oversized OBU can bypass a safety check on platforms where long is 32-bit (e.g., 64-bit Windows), leading to out-of-bounds memory access. Fixed in commit b4c199c5.

[All CVE scores above are CVSS 3.1]

AISLE Doesn’t Complement Legacy SAST. It Replaces It

In one month, AISLE discovered 28 valid security bugs in FFmpeg, along with over 50 CVEs in a variety of other projects, including FreeBSD, OpenSSH, RabbitMQ, and Wireshark. Many of these are bugs a rule-based SAST scanner was never going to surface. AISLE finds them because it reasons about the code rather than matching it against a rule set.

That conclusion is also supported by AISLE’s new benchmark, HoF-Bench, built from 95 public AI-discovered CVEs across 8 repositories. We deliberately chose small, cheap LLMs to see if even relatively weak models could outperform legacy scanners. The result: even these low-powered AIs found more than five times as many CVEs as traditional SAST tools.

For defenders, the message is clear: the frontier is moving down the stack. Vulnerability-finding that used to need a frontier model now runs on cheap ones, which means attackers have it too. Keeping up takes a system built for this, not a rule set retrofitted with AI.

What To Do if You Run FFmpeg

If you run FFmpeg, we recommend implementing the patches:

See What AISLE Finds in Your Code

The same analyzer that found these CVEs in FFmpeg can run against your code wherever it lives: cloud, on-prem, or fully air-gapped. AISLE Snapshot is a one-time audit that delivers verified findings other tools miss, prioritized by business impact. Get your Snapshot.

Our sincere appreciation goes out to the FFmpeg team. All CVEs were found with AISLE’s AI and responsibly disclosed by Joshua Rogers via VulnCheck.