Research Blog

Weekly insights from our cybersecurity research journey

FuzzingBrain-Bench: Measuring Whether LLMs Can Actually Find Real Bugs

Everyone says language models can find vulnerabilities, but no public benchmark measures finding. FuzzingBrain-Bench closes that gap. It ships 68 real C/C++/Java bugs across 40 projects and 17 fault classes, hides the answer, and grades by the program itself on a monotone capability ladder with strict off-target accounting. Inside the corpus, the difficulty tiers, how to run it in five minutes, and what frontier models actually score.

68 Real Bugs 17 Fault Classes Deterministic Oracle Best Blind 71%
Read the Post →

Recent Posts

Harnesses Have Bugs Too: A Four-Principles Audit of 586 OSS-Fuzz Harnesses

Before generating "good" fuzz harnesses, we need a checkable definition of "good". This post offers the Four Principles, then validates them on 586 production OSS-Fuzz harnesses across 70 projects. 53 fix PRs filed, 45 verified or merged, 14 false positives caught at audit, two real upstream bugs surfaced when broken harnesses were repaired.

Series · Part 1

Two Weeks Inside Chromium: 472 Harnesses, 30 Vulnerabilities Filed Upstream

Point the four principles at Google Chromium and its vendor stack (libvpx, libwebp, libaom, openscreen, dawn, pdfium, sqlite, v8, …). Flip the principles into a generation criterion, pair with Logic Group as the semantic-unit slicer, every harness goes through Stage-4 adversarial probing. Two weeks: 472 harnesses, 30 vulnerabilities filed upstream (16 acked so far — 9 confirmed + 7 fixed), and 52 candidates dropped before any could reach a maintainer.

Series · Part 2