TL;DR

Depthfirst autonomous agent located 21 unknown FFmpeg flaws for roughly $1,000. Chrome 149 addressed a record 429 issues. AI tools are generating vulnerabilities quicker than teams can address them.

A security firm deployed its independent AI agent to uncover 21 new vulnerabilities inside FFmpeg, the widely used open-source media library found in nearly all video-related applications. Depthfirst reports the entire process required about $1,000 worth of processing power. Several of these issues had remained undetected in the source code for over two decades.

Shortly afterward Google released Chrome 149 containing fixes for 429 security problems, the highest count in any single browser version. More than 100 of those rank as critical or high impact. Though separate events, both highlight how AI accelerates vulnerability discovery beyond the rate at which humans can resolve them.

The Depthfirst agent examined FFmpeg's approximately 1.5 million lines of C code and generated working proof-of-concept exploits for each of the 21 zero-days. The majority involve heap or stack overflows within parsers and demuxers covering elements such as the TS demuxer and VP9 decoder. One stack overflow in service-description-table handling originated in 2003.

Nine of the issues now have CVE numbers assigned (CVE-2026-39210 through CVE-2026-39218). The remaining ones received upstream fixes without assigned identifiers yet. Depthfirst released the proof-of-concept materials publicly.

FFmpeg has seen prior AI-assisted searches. Google's Big Sleep agent identified several FFmpeg issues the previous year. Anthropic's Mythos model extracted a 16-year-old H.264 problem along with others from FFmpeg at a cost near $10,000. Depthfirst states it achieved similar results at one-tenth the expense.

Chrome 149's large volume of fixes follows a separate path. Google did not link the 429 issues directly to AI. However the company revised its bug bounty program in April due to a surge of AI-produced submissions and now requires researchers to supply brief reproduction steps rather than lengthy reports that AI systems often generate.

The most severe issue, CVE-2026-10881, received a 9.6 CVSS rating. It consists of an out-of-bounds read and write inside the ANGLE graphics engine allowing a specially crafted page to bypass Chrome's sandbox and execute code on the host system. Google awarded $97,000 for that report. Among the 22 critical bugs, 19 originated from internal discovery.

The trend continues across projects. An autonomous tool recently located an authenticated remote code execution problem in Redis that stayed hidden for more than two years. A February analysis demonstrated an AI agent could create working exploits for over half of 100 actual Linux kernel bugs, surpassing conventional fuzzing methods.

The main challenge has changed. Locating these problems has grown inexpensive. Reviewing the reports, delivering the patches, and ensuring deployment remain difficult. Much of this effort still rests with volunteers and a limited group of human reviewers now tasked with matching machine speed. Mozilla addressed 271 Firefox issues uncovered by Mythos in one operation. The focus has moved past whether AI can locate problems to whether anyone can resolve them in time.