AI Agent Uncovers 21 Zero-Days in FFmpeg; Chrome Patches Record 429 Bugs

Two significant developments in security landed within days of each other this week, both underscoring the accelerating impact of artificial intelligence on vulnerability discovery and response. A security startup reported 21 previously unknown vulnerabilities in FFmpeg, the media library embedded in countless applications that process video and audio, all found by an autonomous AI agent. Simultaneously, Google shipped Chrome 149 with patches for 429 security bugs—the most ever in a single Chrome release. While only the FFmpeg bugs were directly attributed to AI, Chrome's record patch count came after Google overhauled its bounty program to handle a flood of AI-generated reports.

AI-Powered Discovery in FFmpeg

The FFmpeg findings come from depthfirst, whose autonomous security agent scanned the project's roughly 1.5 million lines of C code and produced 21 confirmed zero-day vulnerabilities, each with a reproducible proof-of-concept input. According to the company, the entire scanning operation cost approximately $1,000—a remarkably efficient expenditure for such significant findings.

Several of these vulnerabilities had been latent in the codebase for 15 to 20 years. One particularly striking example is a stack overflow in the service-description-table code that dates back to 2003, sitting untouched for 23 years before being discovered by the AI. Most of the discovered vulnerabilities are heap or stack overflows in various parsers and demuxers, spanning components from the TS demuxer to the VP9 decoder.

depthfirst has assigned CVE identifiers to several of these bugs, with CVE-2026-39210 through CVE-2026-39218 already listed, and notes that the remaining vulnerabilities have been fixed but are not yet assigned CVE numbers. The company has published detailed writeups and proof-of-concept code for security researchers and vendors to understand and address these issues.

The significance of these findings cannot be overstated. FFmpeg is one of the most widely reused codebases in software development, serving as the foundation for countless media applications, streaming services, and video processing tools. Its ubiquity means that vulnerabilities in FFmpeg can have cascading effects across numerous products and services.

Chrome's Record Patch Count

In a separate but related development, Chrome 149 addresses 429 vulnerabilities, setting a new record for the most security fixes in a single Chrome release. Over 100 of these are classified as critical or high severity, with the majority involving use-after-free vulnerabilities and insufficient input validation.

The most severe vulnerability, CVE-2026-10881 (CVSS 9.6), is an out-of-bounds read and write in the ANGLE graphics engine that could allow a specially crafted web page to escape Chrome's sandbox and execute arbitrary code on the host system. Google paid $97,000 for this vulnerability, reflecting its critical nature.

Interestingly, the highest-severity vulnerabilities were predominantly discovered internally: of approximately 90 high-severity bugs, only 10 came from outside researchers, and 19 of the 22 critical vulnerabilities were found by Google's own security teams. This pattern suggests that while external researchers continue to contribute valuable findings, internal security teams remain crucial for identifying the most severe vulnerabilities.

AI's Growing Role in Security

The connection between AI and these security developments extends beyond the FFmpeg findings. Google's decision to overhaul its bounty program in April was directly prompted by a flood of AI-generated submissions. The new program now requires concise reproducers rather than the lengthy writeups that AI systems tend to generate.

This isn't the first instance of AI discovering significant vulnerabilities in FFmpeg. Google's Big Sleep agent previously reported a series of FFmpeg bugs, now visible on the project's security page tagged BIGSLEEP. Similarly, Anthropic's Mythos model identified a 16-year-old H.264 flaw and other vulnerabilities in FFmpeg, with three of these issues being addressed in FFmpeg 8.1.

Just days before these announcements, another autonomous tool discovered an authenticated remote code execution vulnerability in Redis that had been present since version 7.2.0, unnoticed for over two years. These findings align with a February study demonstrating that an AI agent could reproduce working proof-of-concepts for more than half of 100 real Linux kernel N-day vulnerabilities, outperforming traditional fuzzing techniques.

Practical Implications for Organizations

The discovery of these vulnerabilities has immediate practical implications for organizations:

For FFmpeg users:

• Update to the latest patched version as soon as it becomes available
• Prioritize systems that process untrusted RTSP or AV1-over-RTP streams
• Remember that FFmpeg is widely bundled in media pipelines, Python wheels, container images, and appliances
• Don't limit patching to system packages—embedded copies in third-party software also need updating

For Chrome users:

• Update to Chrome 149.0.7827.53 on Linux or 149.0.7827.53/54 on Windows and macOS
• Confirm that auto-updates have been applied
• Consider the increased patch frequency as a reason to keep browsers updated more diligently

The Challenge of Scale

While AI has dramatically accelerated vulnerability discovery, it has also created new challenges for security teams. Finding bugs has become significantly cheaper and faster, but the process of triaging reports, developing fixes, and ensuring widespread deployment has not kept pace.

Much of this work still falls to volunteers and a thin layer of human security professionals who are now expected to keep pace with machine-generated findings. The traditional security lifecycle—from discovery to patch to deployment—is being compressed, requiring organizations to adapt their processes.

Security teams are now exploring strategies like shorter patch cycles, more aggressive auto-deployment of updates, and treating dependency updates that include CVE fixes as security work rather than routine maintenance. However, these approaches require organizational changes that can be difficult to implement quickly.

The Future of AI in Security

These developments represent just the beginning of AI's impact on cybersecurity. As AI systems become more sophisticated and widely adopted, we can expect:

• Even faster discovery of vulnerabilities across a broader range of software
• More automated vulnerability assessment and remediation
• New challenges in distinguishing between valid and false-positive findings
• Potential shifts in the economics of security research and bug bounty programs

For organizations, the key takeaway is clear: the pace of vulnerability discovery is accelerating, and traditional security practices may no longer be sufficient. Proactive measures, including automated patching, dependency scanning, and streamlined update processes, are becoming essential components of a robust security strategy.