AI Agent Discovers 21 FFmpeg Zero-Days: CVE-2026-39210-39218

An autonomous AI security agent has uncovered 21 zero-day vulnerabilities in FFmpeg, the ubiquitous media processing library that powers browsers, streaming platforms, and countless applications worldwide. The findings mark a significant milestone in AI-driven vulnerability discovery, with some bugs remaining latent in the codebase for 15 to 23 years.

What Happened: AI Agent Discovers FFmpeg Zero-Day Cluster

Security startup depthfirst announced on June 4, 2026, that its production autonomous security agent had successfully identified and verified 21 previously unknown vulnerabilities in FFmpeg. The discovery comes just days after Google shipped Chrome 149 with a record-breaking 429 security patches, highlighting the accelerating pace of AI-augmented security research.

The depthfirst agent analyzed FFmpeg’s approximately 1.5 million lines of C code and produced concrete, reproducible proof-of-concept inputs for each vulnerability. The company reports the entire operation cost approximately $1,000, demonstrating the cost-efficiency of autonomous security analysis compared to traditional manual audits.

Several of the discovered vulnerabilities had been dormant in the codebase for decades. One particularly severe stack overflow in the service-description-table code dated back to 2003 and had remained undetected for 23 years despite extensive fuzzing and manual review efforts.

Technical Details of the FFmpeg Vulnerabilities

The 21 vulnerabilities primarily consist of heap and stack buffer overflows located in various parsers and demuxers throughout FFmpeg’s extensive codec library. The findings span multiple components, from the TS demuxer to the VP9 decoder, affecting how FFmpeg processes malformed media files.

Vulnerability Categories

• Heap buffer overflows: Memory corruption in dynamic allocation regions during media parsing
• Stack buffer overflows: Overflows in fixed-size buffers on the call stack
• Parser vulnerabilities: Flaws in TS demuxer, VP9 decoder, and service-description-table handling

Affected Components

The vulnerabilities impact multiple FFmpeg components responsible for processing various media formats. Consequently, any application using FFmpeg to handle untrusted media files from external sources faces potential exploitation risk.

Nine of the vulnerabilities have already received CVE identifiers:

• CVE-2026-39210
• CVE-2026-39211
• CVE-2026-39212
• CVE-2026-39213
• CVE-2026-39214
• CVE-2026-39215
• CVE-2026-39216
• CVE-2026-39217
• CVE-2026-39218

Depthfirst has published proof-of-concept code demonstrating the vulnerabilities and developed an RCE exploit primitive to confirm exploitability.

Business and Operational Impact

The widespread deployment of FFmpeg creates significant exposure across multiple sectors. Moreover, the nature of these vulnerabilities enables zero-click attacks through crafted media files.

Affected Industries

• Web browsers: Chrome, Firefox, and Edge all rely on FFmpeg components
• Streaming platforms: Media processing pipelines at risk
• Enterprise software: Video conferencing and content management systems
• Mobile applications: iOS and Android apps using FFmpeg libraries
• IoT devices: Smart TVs and media players with FFmpeg integration

Risk Assessment

The vulnerabilities present varying levels of risk depending on deployment context. Applications that process untrusted media files from the internet face the highest exposure. Furthermore, the long dormancy of some bugs suggests traditional security testing methods may not catch similar issues in other widely-deployed libraries.

Mitigation and Recommendations

Organizations should take immediate steps to assess their FFmpeg exposure and apply available protections.

Immediate Actions for Defenders

1. Inventory FFmpeg usage: Identify all applications and systems using FFmpeg libraries
2. Apply latest updates: Update to FFmpeg version 7.1.2 or later containing fixes
3. Implement input validation: Sanitize and validate all media files before processing
4. Deploy sandboxing: Isolate media processing in restricted environments
5. Monitor for anomalies: Watch for unexpected crashes in media handling components

Long-term Security Strategies

Organizations should consider adopting AI-augmented security scanning for their own codebases. The cost efficiency demonstrated by depthfirst suggests autonomous agents may become standard tools for vulnerability discovery. Additionally, security teams should evaluate their fuzzing programs to ensure coverage gaps are addressed.

Bottom line: With FFmpeg embedded in billions of devices and applications, these AI-discovered vulnerabilities represent a wake-up call for security teams. Immediate patching and exposure assessment are critical, while the broader implications suggest AI agents will increasingly complement human security researchers in finding long-dormant bugs.

Incident Summary

CVE IDs:	CVE-2026-39210 through CVE-2026-39218 (9 assigned), plus 12 additional findings
Affected Systems:	FFmpeg media processing library, versions prior to 7.1.2
Disclosure Date:	June 4, 2026
Patch Status:	Patches available in FFmpeg 7.1.2
Discovery Method:	Autonomous AI security agent (depthfirst)
Discovery Cost:	Approximately $1,000 USD

References

1. Depthfirst Research, “21 Zero-Days in FFmpeg,” June 4, 2026, https://depthfirst.com/research/21-zero-days-in-ffmpeg, accessed June 6, 2026.
2. The Hacker News, “AI Agent Uncovers 21 Zero-Days in FFmpeg; Chrome Patches Record 429 Bugs,” June 6, 2026, https://thehackernews.com/2026/06/ai-agent-uncovers-21-zero-days-in.html, accessed June 6, 2026.
3. FFmpeg Security Page, https://ffmpeg.org/security.html, accessed June 6, 2026.
4. Google Big Sleep Team, “Finding Zero-Day Vulnerabilities with Large Language Models,” https://goo.gle/bigsleep, accessed June 6, 2026.
5. Anthropic Research, “Mythos Preview,” April 2026, https://red.anthropic.com/2026/mythos-preview/, accessed June 6, 2026.
6. GitHub Repository, “FFmpeg DF Vulnerability Proof-of-Concept,” https://github.com/DepthFirstDisclosures/ffmpeg-dfvuln127, accessed June 6, 2026.