๐The Next Web (TNW)โขStalecollected in 59m
AI agent discovers 21 zero-day vulnerabilities in FFmpeg
๐กAI agents are now finding 20-year-old bugs for $1,000. See how autonomous security research is changing.
โก 30-Second TL;DR
What Changed
Autonomous AI agent found 21 zero-day bugs in FFmpeg.
Why It Matters
This demonstrates the high efficiency of AI in automated vulnerability research. It highlights the potential for AI to secure legacy codebases that are too large for manual review.
What To Do Next
Evaluate your dependency tree for legacy libraries like FFmpeg and ensure they are updated to the latest patched versions.
Who should care:Researchers & Academics
Key Points
- โขAutonomous AI agent found 21 zero-day bugs in FFmpeg.
- โขThe discovery process cost approximately $1,000 in compute resources.
- โขSome vulnerabilities had persisted in the codebase for over 20 years.
๐ง Deep Insight
Web-grounded analysis with 9 cited sources.
๐ Enhanced Key Takeaways
- โขThe security startup responsible for the discovery is named depthfirst, and their autonomous security agent scanned approximately 1.5 million lines of C code in the FFmpeg project.
- โขThe identified vulnerabilities primarily consist of heap or stack overflows found in FFmpeg's parsers and demuxers, impacting components such as the TS demuxer and VP9 decoder.
- โขNine of the 21 zero-day vulnerabilities discovered by depthfirst's AI agent have already been assigned CVE identifiers, ranging from CVE-2026-39210 to CVE-2026-39218.
- โขOne particularly old vulnerability, a stack overflow in the service-description-table code, dates back to 2003, meaning it remained undetected for 23 years.
- โขdepthfirst has published proof-of-concept (PoC) inputs for the discovered vulnerabilities, with the remaining bugs fixed upstream but not yet numbered.
๐ Competitor Analysisโธ Show
| Feature/Company | depthfirst AI Agent | Google's Big Sleep Agent | Anthropic's Mythos/Claude Model |
|---|---|---|---|
| Target Software | FFmpeg | FFmpeg | FFmpeg, OpenBSD, Linux kernel, web browsers, various systems |
| Vulnerabilities Found (FFmpeg) | 21 zero-days | Multiple bugs (reported in 2025) | 16-year-old H.264 flaw, 500+ zero-days across various systems |
| Compute Cost (FFmpeg) | ~$1,000 | Not specified | ~$10,000 for FFmpeg flaw |
| Oldest Bug Found | 23 years (stack overflow in service-description-table) | Not specified | 16 years (H.264 flaw in FFmpeg), 27 years (OpenBSD bug) |
| Claimed Efficiency | Comparable work to Anthropic's Mythos at 1/10th the cost | Not specified | Scored 83.1% on CyberGym (vulnerability reproduction) vs 66.6% for Opus 4.6 |
| Availability | Commercial service | Internal Google agent | Unreleased (given to partners for critical infrastructure) |
๐ ๏ธ Technical Deep Dive
- Code Scanning: depthfirst's AI agent scanned approximately 1.5 million lines of C code in FFmpeg.
- Hypothesis Generation & Testing: Anthropic's Claude Code and Mythos Preview models read source code, formulate hypotheses about potential vulnerabilities, and then execute the software to test these hypotheses.
- Debugging & Iteration: The AI agents can add debug logic or utilize debuggers as needed, iterating through the process until a bug is identified or the search is abandoned.
- Isolated Environment: Anthropic's testing methodology involves launching isolated containers without internet access, containing only the software under examination and its source code.
- Hybrid Approach: Other approaches, like ZeroPath's, combine deep program analysis with adversarial AI agents for vulnerability validation.
- General AI in Cybersecurity: AI in cybersecurity broadly leverages machine learning to analyze network patterns, detect anomalies, and pinpoint system weaknesses. Generative AI is also used to correlate and analyze information for reporting.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
AI-driven vulnerability discovery will significantly increase the volume and speed of reported zero-day vulnerabilities.
The demonstrated efficiency and low compute cost of AI agents in finding long-standing bugs suggest a future where more vulnerabilities are found faster than human teams can manage, as indicated by Google's overhaul of its bounty program due to a flood of AI-generated reports.
The cost of discovering zero-day vulnerabilities will drastically decrease, democratizing access to advanced bug-hunting capabilities.
The $1,000 compute cost for 21 zero-days by depthfirst and Anthropic's $10,000 for a significant flaw in FFmpeg show a substantial reduction in the financial barrier to entry for finding critical vulnerabilities, potentially enabling more actors (both benevolent and malicious) to engage in such activities.
Software development and patching processes will need to adapt to a continuous, AI-driven vulnerability disclosure landscape.
The rapid discovery of numerous, long-standing bugs by AI agents necessitates faster and more automated patching mechanisms, and a shift towards proactive security integration throughout the software development lifecycle to keep pace with AI's ability to find flaws.
โณ Timeline
2016
DARPA Cyber Grand Challenge showcases AI systems autonomously finding and patching vulnerabilities.
2023-Early
AIxCC researchers demonstrate first practical applications of LLM-powered vulnerability detection.
2024-07
ZeroPath begins combining deep program analysis with adversarial AI agents for validation, uncovering critical vulnerabilities.
2025
Google's Big Sleep agent reports FFmpeg bugs.
2026-04-07
Anthropic announces Project Glasswing and Claude Mythos Preview, an AI model finding zero-days in major OS and browsers, including FFmpeg.
2026-06-06
depthfirst's AI agent discovers 21 zero-day vulnerabilities in FFmpeg.
๐ Sources (9)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Next Web (TNW) โ