Fuzzilli vs BigSleep: 37 and 45 Vulnerabilities Found in V8 Over Six Months with 25% Overlap

Google's Project Zero compared two automated vulnerability discovery tools targeting the V8 JavaScript engine in Chrome. The generative fuzzer Fuzzilli identified 37 actionable vulnerabilities over six months, while the LLM-based agent BigSleep found 45. Only about 25% of the vulnerabilities overlapped between the two systems. Additional experiments on historical bugs from 2023 onward showed a theoretical overlap of around 60%. The analysis indicates that fuzzing and LLM agents complement each other rather than one replacing the other. Researchers recommend using both approaches together to maximize coverage of complex bugs in well-tested software like V8.

HabrOctober 28, 2026Translated from Russian

Read at source

Google's Project Zero team conducted a detailed comparison of two automated tools used to hunt vulnerabilities in V8, the JavaScript and WebAssembly engine powering Chrome. The study pitted the generative JavaScript fuzzer Fuzzilli against BigSleep, an agentic system powered by large language models. Over a six-month period from April to September 2026, BigSleep discovered more vulnerabilities than Fuzzilli (45 versus 37), yet only roughly 25% of the bugs overlapped. Experiments on older bugs suggested the true intersection could reach approximately 60%.

Production Statistics

Both systems operate in production and report crashes through Clusterfuzz, which performs initial triage and creates bug reports in the Chromium tracker. Fuzzilli runs continuously across many instances and typically picks up new code changes within hours. BigSleep runs once per day, reviewing recent commits and generating proof-of-concept test cases for discovered issues.

Raw reports showed Fuzzilli producing 316 bugs and BigSleep producing 441. After filtering for vulnerabilities only and then for actionable reports (those not marked duplicate or WAI and linked to at least one fixing commit), the numbers became 37 for Fuzzilli and 45 for BigSleep. This represents a near tie with a slight edge for the LLM-based agent.

Overlap Analysis

Researchers measured overlap in two ways. First, they examined the bug tracker for reports filed by one system that were marked duplicates of reports from the other, including indirect duplicates resolved through Clusterfuzz signatures and fixing commits. For actionable vulnerabilities, the overlap reached approximately 25% (20 out of 81), with slightly more than half first discovered by Fuzzilli.

In a second experiment, BigSleep was tasked with rediscovering 112 historical Fuzzilli bugs from 2023 to mid-2026. The agent successfully reproduced the original bug in 62% of cases and found a different (already fixed) bug in another 12%. No statistically significant memorization effect was observed when comparing success rates across different time periods relative to the model's knowledge cutoff.

Bug Complexity and Components

Analysis of rediscovery rates by bug complexity (measured by time taken for Fuzzilli to find the issue) showed that harder-to-find bugs for the fuzzer were also harder for the LLM agent. Overlap was notably lower for bugs in compiler components, where Fuzzilli has specialized strengths.

The study concludes that both techniques remain valuable. A substantial portion of vulnerabilities was found exclusively by one method or the other. The rapid progress of LLM agents is encouraging, yet fuzzing and agentic analysis currently complement each other. The authors recommend that software projects apply both fuzzing and LLM-based analysis to reduce the number of vulnerabilities reaching end users.

Gemini Project Zero Fuzzilli BigSleep Clusterfuzz

ShareX Facebook LinkedIn Reddit

Meridian digestThe week's most consequential security stories, distilled into one briefing.See the digest

More in Vulnerabilities & Exploits

Vulnerabilities & Exploits

Telegram for Android Patches Voice Call Vulnerability Enabling Code Execution

Security researchers at NebuSec disclosed a vulnerability in Telegram for Android that allowed remote code execution after a victim answered a voice call. The flaw affected versions from 8.6.2 through 12.10.3 and was fixed in release 12.10.4 at the end of September. The issue stemmed from improper handling of MediaState service messages that carry microphone, camera, and video rotation data during calls. When a required field was missing, the application accessed invalid memory. NebuSec discovered the bug using its Code Security product running on a GPT-5.4 model and confirmed exploitation on ARM64 Android devices for personal calls. A demonstration showed the string “HACKED BY NEBUSEC” appearing roughly 30 seconds after the call was accepted. The vulnerability is unrelated to the earlier animated-sticker attack vector.

AntiMalwareOct 29, 2026Translated from Russian

Vulnerabilities & Exploits

Yandex Patches Millions of Stations After eMudhra Certificate Keys Compromised

Yandex has automatically deployed security updates to millions of its smart speakers and partner televisions running YaOS between October 21 and 23. The patches address compromised intermediate certificates issued by eMudhra whose private keys consisted of trivial values from 1 to 255. Attackers who obtained these keys could generate fraudulent certificates and perform man-in-the-middle attacks against device HTTPS traffic. Enthusiasts demonstrated that local network control combined with the certificates allowed replacement of server configuration, activation of ADB, and acquisition of root access on the devices. The update removes trust in the affected certificates, overwrites compromised system files, and disables the debug interface. A factory reset restores the vulnerable firmware, so Yandex advises against resets without a compelling reason.

AntiMalwareOct 29, 2026Translated from Russian

Vulnerabilities & Exploits

Cisco ASA and FTD Face DoS Vulnerability in SSL VPN Already Exploited in the Wild

Cisco Systems has disclosed a denial-of-service vulnerability affecting its Cisco Secure Firewall ASA and Cisco Secure Firewall FTD products. The flaw, tracked as CVE-2026-76418, resides in the SSL VPN remote access feature and carries a CVSS v3.1 base score of 8.6, rated High severity. Exploitation occurs when secondary authentication is configured on tunnel groups, allowing an attacker to send a crafted request that forces a device reboot without requiring authentication. Cisco confirmed active exploitation of the issue in October 2026. Security updates addressing the vulnerability are now available for both affected product lines. The advisory recommends immediate application of the patches to prevent service disruption.

Security NEXTOct 29, 2026Translated from Japanese