BigSleep Finds 45 V8 Vulnerabilities to Fuzzilli's 37 in Six Months, With Only 25% Overlap

Project Zero reports that BigSleep, an LLM-based bug finding system, found more vulnerabilities in Chrome's V8 JavaScript engine than the Fuzzilli fuzzer over the past six months: 45 versus 37. Only 25% of the specific bugs reported were shared between the two systems.

What was compared

Fuzzilli is a generative JavaScript engine fuzzer. BigSleep is an agentic, LLM-based bug finding system. Both run in production at Google and automatically look for bugs in newly added V8 code, which Project Zero says makes them fairly directly comparable.

• Fuzzilli runs continuously across many instances. The first instances usually pick up new changes within a few hours.
• BigSleep currently runs once per day. It is prompted to review new commits and write proof-of-concept testcases for any bugs it discovers.
• Both report bugs through Clusterfuzz, which does basic crash triaging and then files reports in Chrome's issue tracker.

The analysis covers April to September 2026. Project Zero limited the window to those six months because the automatic BigSleep pipeline only became active in April 2026. The comparison is restricted to V8 as a target. Project Zero leaves other targets, fuzzers and agentic systems to future work.

Production statistics

Raw bug report counts from the Chromium issue tracker hotlists for the period:

Metric	Fuzzilli	BigSleep
All bug reports	316	441
Reports typed as vulnerability	178	95

Project Zero cautions that the raw counts are misleading. BigSleep initially found many configuration and integration issues in V8, such as an unsafe flag blocking mechanism that did not work properly. These are legitimate issues but carry little real-world security impact. Duplicates and unactionable reports, such as bugs closed as working as intended, also distort the totals.

Project Zero also describes a Fuzzilli report that Clusterfuzz filed and closed on the same day with no developer interaction. It was most likely a duplicate that Clusterfuzz missed because the crash signature differed slightly. Because of cases like this, filtering on fixed or verified status is not straightforward. Project Zero instead defines "actionable" bugs as those not marked duplicate or working as intended and referenced by at least one fix commit, then filters to vulnerability-type bugs. On that basis, BigSleep found 45 vulnerabilities and Fuzzilli 37.

Overlap

In the six-month comparison, 25% of the specific bugs reported overlapped. Project Zero says this methodology tends to underestimate overlap, so it ran dedicated experiments to test theoretical overlap for vulnerabilities reported since 2023. Those experiments estimated overlap at roughly 60%.

Project Zero states that regardless of which value is accurate, a substantial portion of vulnerabilities were discoverable by only LLMs or only fuzzing. The post also examines the "shape" of bugs found by one system but not the other.

What this means

The headline count favors BigSleep, but Project Zero's own data shows the picture depends on how bugs are counted. BigSleep leads on total reports (441 versus 316), while Fuzzilli leads on reports typed as vulnerabilities (178 versus 95). The 45 versus 37 figure comes from the stricter definition of actionable vulnerabilities.

The overlap figures, 25% observed and roughly 60% estimated, both leave a substantial share of bugs that only one approach found. Project Zero describes it as extremely promising that LLMs can find so many vulnerabilities so quickly, but concludes that fuzzing and LLMs remain complementary. It recommends that vendors run both fuzzing and agentic analysis of their code to maximize the number of vulnerabilities reported before they impact users.

The findings apply to V8, one target with a mature fuzzing setup, and to these two systems only.