AISLE Finds 21 Security Issues in FFmpeg, Including 6 New CVEs

FFmpeg acts as the hidden engine powering much of the internet's video processing, from archaic formats to the bleeding edge. Due to its central role, it’s been reviewed extensively by humans and machines ever since it launched in 2000, making new discoveries difficult. Yet when we used AISLE to analyze the codebase, we found 21 security issues, 6 of which received CVE designations, including 3 code execution primitives and 2 out-of-bounds read issues.

These findings demonstrate the power of multi-model AI systems, even on mature, hardened codebases.

AISLE: System Over Model

When Google and Anthropic used AI to scan FFmpeg earlier this year, the story revolved around the compute power these frontier labs can bring to bear on cybersecurity tasks. At AISLE, however, we have long argued that security capability is not a function of raw model size, but of the way that a well-engineered AI system matches model strengths to the rigors of each distinct cybersecurity task. 

As AISLE co-founder and Chief Scientist Stanislav Fort demonstrated in April, small, open-weight models often outperform frontier giants at discrete, well-defined cybersecurity tasks, such as broad-spectrum scanning, false positive identification, and exploit construction. By implication, a multi-model system can selectively deploy models to complete the tasks they are best suited for, delivering better results for far lower compute.

And of course, that translates into far lower costs. In Fort’s research, a tiny open-weight model that runs 600x cheaper than Mythos matched it discovery by discovery on Anthropic’s flagship FreeBSD findings. 

Those cost savings are not simply financial metrics. When defenders are no longer constrained by compute budgets, they can run AI code analysis continuously on every single line of code, including the auxiliary systems where vulnerabilities often hide unnoticed for years. Thus, multi-model systems like AISLE (finally) align financial incentives with cybersecurity outcomes.

The results are clear. At the time of writing, AISLE leads all AI cybersecurity companies with 279 CVEs (30 designated as critical), 21 MITRE-Top 25 classes, and 88 distinct CWE IDs. To see what this detection engine finds in your codebase, get a one-time code audit, AISLE Snapshot, wherever your code lives: air-gapped, on-prem, or in the cloud.

What the FFmpeg Findings Tell Us About AI Cybersecurity

As a vital, mature, and elegant codebase, FFmpeg is an ideal proving ground for our thesis. Hardened over 25+ years, during which time it’s been subjected to rigorous analysis by the latest tools and methods, this multimedia framework is both well-regarded and ubiquitous.

When we used AISLE to analyze the codebase, we found 21 valid findings, including 6 that resulted in CVEs. Perhaps the most alarming of these were the three code execution primitives, which due to the fundamental role FFmpeg plays in video streaming, would potentially enable remote code execution on any system that processes arbitrary media files, such as YouTube or Vimeo. Each of these high severity vulnerabilities received an 8.7 CVSS score:

1. CVE-2026-64830

This heap buffer overflow allows attackers to craft a subtitle file with excessive distinct stream IDs to potentially trigger unbounded writes and achieve arbitrary code execution in any application that uses FFmpeg’s VobSub demuxer.

2. CVE-2026-64831

This stack buffer overflow allows remote attackers to overwrite return addresses and adjacent stack frames, potentially overflowing stack-allocated arrays and achieving arbitrary code execution.

3. CVE-2026-64832

This double-free vulnerability allows attackers to trigger memory corruption via a crafted video file, and when no decoder surfaces are left, the error path frees memory while the calling layer frees the same frame description, resulting in a double-free. 

The remaining findings are not trivial. In fact, two of them, CVE-2026-64833 and CVE-2026-64835, are out-of-bounds reads that would allow attackers to steal memory across expected boundaries. Finally, CVE-2026-64834 allows FFmpeg to run forever, which would have disparate impacts on different systems. 

In addition to these 6 CVEs, AISLE found 15 additional valid issues. Taken together, these discoveries are a clear sign that cybersecurity capability is jagged, and defenders need an intelligent system that puts the best-performing models to the task, not one that locks them into one specific (and costly) LLM. 

What to Do if You Run FFmpeg

The following bugs have been fixed. If you run FFmpeg, we recommend implementing the patches:

1. CVE-2026-64830 (VobSub demuxer heap overflow)
2. CVE-2026-64831 (Vulkan HEVC stack overflow)
3. CVE-2026-64832 (NVDEC double-free)
4. CVE-2026-64833 (S/PDIF muxer OOB read)
5. CVE-2026-64834 (RTP/ASF infinite loop)
6. CVE-2026-64835 (ADX decoder OOB)

Aligning Model Capability and Cybersecurity Need With AISLE

AISLE is engineered to align cost incentives, security capability, and compute power to deliver the most performant AI security platform into any environment. Want to see what it finds in your codebase? AISLE Snapshot is a one-time code audit that works wherever your code lives: air-gapped, on-prem, or in the cloud. Get your Snapshot.

AISLE follows responsible disclosure practices, and would like to thank the FFmpeg team for their professionalism and collaboration. These findings were reported by Pavel Kohout of the AISLE Research Team.