What changed
On October 27, 2026, Google Project Zero published a head-to-head analysis of BigSleep, its LLM-driven bug-finding system, and Fuzzilli, a specialised JavaScript-engine fuzzer. Both report into Chrome's Clusterfuzz/issue tracker and were compared for April–September 2026. After removing duplicate, non-actionable and non-vulnerability reports, BigSleep found 45 actionable V8 vulnerabilities versus 37 for Fuzzilli. Only 20 of 81 distinct vulnerabilities were found by both (roughly 25% overlap). A separate controlled exercise on 112 historical Fuzzilli vulnerabilities found BigSleep could rediscover about 62% with 25 attempts per commit, increasing to around 70% with additional prompting and starting-point variants.
Why it matters
Engineering teams increasingly ask whether AI security review can replace traditional fuzzing. This unusually concrete production comparison suggests the opposite: each method catches bugs the other misses, especially when compiler internals, tricky language semantics, runtime flags or multi-stage interactions matter. The actionable practice is to add agentic code analysis alongside mature fuzzing, not withdraw fuzzing coverage based on headline AI bug counts. The study is Google-specific, does not measure independent tool costs and cannot prove the same performance on other projects.
The 45-versus-37 figure is filtered, not raw alerts
Project Zero began with 441 BigSleep and 316 Fuzzilli raw reports, then excluded duplicates, unhelpful reports and non-security findings. The final comparison counts 45 versus 37 actionable vulnerability reports. Because the pipelines operate at different cadences—BigSleep reviews new commits daily, while Fuzzilli fuzzes continuously—speed and opportunity to rediscover already-fixed bugs affect observed overlap.
Production overlap was about 25%; the controlled test was different
Only 20 of 81 distinct actionable vulnerabilities were shared in production. Project Zero then reconstructed 112 older fuzzer-found bugs as commit-review challenges and ran 25 BigSleep attempts per starting point. The agent rediscovered about 62% of the target bugs and found different already-fixed bugs on some others. This controlled result estimates what an agent could find with a selected starting point, not the overlap expected in day-to-day production.
Different techniques expose different failure modes
BigSleep can reason through complex language control flow and edge-case values that are difficult to generate by random mutation. Fuzzilli remains strong at compiler interactions, nonlocal failures and coverage-guided exploration, and is not biased by a particular prompt or commit selection. The experiment also found that bugs hard for fuzzers tended to be hard for the agent, so the two approaches are complementary rather than neatly partitioned.
Operational implications and what remains unknown
Security teams should retain their existing fuzzing pipeline and trial agentic review on recent commits, recording unique actionable vulnerabilities, time to discovery, reproducibility and human triage burden. Project Zero could not compare cost per actionable bug fairly because neither pipeline was tuned for that metric. Its findings are limited to one mature engine and particular tool/model versions; they should not be presented as a universal ranking of AI versus fuzzing.