Fuzzilli vs BigSleep: 37 and 45 Vulnerabilities Found in V8 Over Six Months with 25% Overlap
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.
Related articles
Zero-Day Vulnerabilities: The Invisible Threats Lurking in Modern Software
Zero-Day vulnerabilities represent one of the most challenging problems in cybersecurity because developers have no time to prepare defenses before exploitation begins. These flaws exist in software components without the knowledge of vendors, leaving antivirus tools and existing security mechanisms unable to detect attacks. Researchers discover such issues through techniques including fuzzing, reverse engineering, and static analysis. High-profile examples such as Heartbleed in OpenSSL, EternalBlue in Windows SMB, and Log4Shell in the Log4j library illustrate how single errors can affect millions of systems worldwide. Bug Bounty programs now allow companies to harness independent researchers to find problems before criminals exploit them. Organizations can reduce risk through regular updates, least-privilege principles, multi-factor authentication, and layered defenses, though complete elimination of zero-day risk remains impossible.
11-Year-Old Linux Kernel Race Condition NoisyTCP (CVE-2026-99288) Enables Remote DoS on Nginx, Apache and SSH Servers
Security researcher Hyunwoo Kim disclosed NoisyTCP (CVE-2026-99288), an 11-year-old race condition in the Linux kernel networking stack that allows unauthenticated remote attackers to trigger kernel panics and crash servers. The flaw stems from concurrent access to the ehash table during TCP socket lookup, where one CPU context may dereference a freed request_sock or full socket while another reuses the memory via SYN cookies. Exploitation requires multiple RX queues or RPS enabled and succeeds reliably on virtualized environments within seconds to minutes. Affected systems include nearly all distributions running kernels from 4.4 onward until the fix landed in November 2026. Temporary mitigations involve disabling SYN cookies or restricting traffic to a single RX queue, though both reduce resilience or performance. The researcher published full technical details on GitHub after coordinated disclosure.
Check Point Gateways Hit by PHP2Shell Exploits Installing TSGA Backdoor via Chained CVEs
An unidentified threat actor has been actively exploiting known vulnerabilities in unpatched Check Point Security Gateways to deploy the TSGA backdoor. The campaign began on 30 October 2026 against Israeli organizations and expanded on 2 November to targets across multiple countries and industries. Initial access leveraged CVE-2026-109618 and CVE-2026-109619 in exposed Mobile Access portals, followed by local privilege escalation through the previously unknown CVE-2026-123039 in Gaia OS. Attackers then installed PHP web shells and the TSGA backdoor, which provides encrypted remote access, configuration theft, persistence mechanisms, and log tampering via the embedded TSProxy component. In the first wave, operators pivoted to VMware ESXi hosts and deployed the destructive wiper destroy_v6. While tooling overlaps with the China-linked Armory Castle group’s PhiliKit framework, the use of a wiper and references to the Israel-Palestine conflict leave attribution open to three scenarios.
Apache Shiro Authentication Bypass Vulnerability Disclosed, Fixed in August Update
A critical authentication bypass vulnerability has been identified in the Java security framework Apache Shiro under specific LDAP caching configurations. The flaw, tracked as CVE-2026-129349, affects versions prior to 3.0.0 and carries a CVSSv4.0 base score of 9.1, rated Critical. Once a valid user authenticates successfully, subsequent login attempts using the cache accept any password for that username, enabling impersonation attacks. The Apache Shiro development team disclosed the issue on November 14, 2026, noting that the vulnerability was already resolved in version 3.0.1 released in August 2026. Version 3.0.2 was made available on November 8, 2026, and users are urged to upgrade immediately to the latest release.