DepthFirst agent discovers 21 FFmpeg zero-days
DepthFirst's autonomous security agent has discovered 21 zero-day vulnerabilities in FFmpeg, including flaws present for over a decade. The findings highlight the efficiency of agentic security testing, identifying a critical remote code execution flaw in the AV1 RTP depacketizer at a fraction of the cost of previous audits.
While LLM-powered coding assistants typically focus on interactive generation, specialized security agents utilizing parallel hypothesis testing and execution feedback can discover critical vulnerabilities overlooked by manual audits and traditional fuzzers.
- –**Agentic Efficiency**: The autonomous agent found 21 zero-days at a cost of ~$1k, compared to Anthropic's $10k Mythos scan, proving the efficiency of targeted execution validation over raw scale.
- –**Deep Latency**: Several discovered bugs (e.g., DFVULN-122 in the RTP MPEG-4 depacketizer) have been present since 2003-2010, demonstrating that legacy code remains a major source of hidden risk even in heavily audited projects.
- –**Practical Exploitability**: The AV1 RTP depacketizer vulnerability (DFVULN-127) offers a clean RCE primitive by allowing the attacker to poison write cursors and hijack AVBuffer.free without corrupting refcounts.
DISCOVERED
2h ago
2026-06-13
PUBLISHED
4h ago
2026-06-12
RELEVANCE
AUTHOR
redbell