BUGBANK’s Elliot Tops Two Categories in UC Berkeley’ CyberGym Benchmark, Including Single-Model and Test-Time Memory
SHANGHAI, CN / ACCESS Newswire / November 19, 2026 / The CyberGym Leaderboard (Level 1), a cybersecurity vulnerability
Press Release Disclaimer: This is a press release distributed through the XPR Media network. It has not been independently verified by our newsroom.
![]()
SHANGHAI, CN / ACCESS Newswire / November 19, 2026 / The CyberGym Leaderboard (Level 1), a cybersecurity vulnerability discovery benchmark maintained by the Center for Responsible, Decentralized Intelligence (RDI) at the University of California, Berkeley, was updated on November 10. The “AI white-hat hacker” system elliot-v1, tested by the Elliot team under Shanghai Moule Network Technology Co., Ltd. (BUGBANK), achieved an overall success rate of 97.28% with one trial per task (trials: 1), entering the official top tier and ranking No. 7 overall on the leaderboard.
In specific categories, the system ranked first in both the “single self-developed security model” and “test-time self-evolving memory (test-time mem.)” dimensions.
CyberGym: A Vulnerability Discovery Benchmark Based on Real-World Industrial Code
CyberGym is an AI security evaluation benchmark proposed by UC Berkeley. According to public information, the benchmark has been cited in the International AI Safety Report 2026, as one of four global AI safety benchmarks tracked by the report.
The benchmark uses real-world industrial-grade codebases, covering 188 major large-scale open-source projects and 1,507 historical high-severity vulnerabilities from ARVO and OSS-Fuzz. During testing, participating systems are not provided with any solution hints; they must rely entirely on autonomous reasoning to locate vulnerabilities in source code, synthesize attack inputs, and complete crash verification. Based on public scores, elliot-v1 demonstrated consistent vulnerability reproduction performance on both vulnerability sets, with its overall single-trial success rate surpassing participating systems from multiple international technology companies.
Two Differentiated Technical Approaches
Publicly available technical configurations show that elliot-v1, which ranks No. 7 overall and is equipped with the self-developed security model elliot-rdt-v1, differs from front-running systems along two engineering dimensions:
Underlying Model Dimension: No. 1 Among Systems Using a Single Self-Developed Security Model.
The top six systems on the leaderboard generally employ multi-model collaborative scheduling or are built directly on publicly available general-purpose large models. From a practical standpoint, multi-model concurrent scheduling entails higher computational costs, constraining enterprise-scale deployment; meanwhile, general-purpose large models are limited by built-in safety alignment policies and often require unconventional techniques such as “jailbreaking” to be used for vulnerability attack testing, making it difficult to meet compliance requirements in commercial applications. elliot-v1 competes independently with a single vertical security model, supports private self-hosting, and achieves the highest score among comparable proprietary single-model architectures.
Agent Memory Dimension: No. 1 in Test-Time Self-Evolving Memory.
Real-world cyber offense and defense are typically long-duration, cross-project continuous tasks rather than isolated single-task tests. When handling long-sequence tasks, large models are susceptible to context window and attention mechanism constraints, leading to memory bloat, semantic misalignment, and retrieval drift. The test-time mem. tag group established by CyberGym is designed to examine whether an agent can extract vulnerability patterns and construction paradigms in real time during the testing process and transfer and reuse them across tasks. elliot-v1 ranks first in this official category.
Steve, CEO of Shanghai Moule Technology (BUGBANK), stated: “The competition in vulnerability discovery ultimately is not about whose model is larger, but about whose system can continually develop better judgment in sustained, real-world offensive and defensive operations-this is the fundamental reason we chose the path of a single model plus self-evolving memory.”
A technical expert noted that the above two metrics address the core issues in the enterprise-level deployment of AI vulnerability discovery-how to use computational resources cost-effectively while ensuring technical compliance and stable long-term system operation.
From a Crowdsourced Testing Platform to a Self-Developed Security Model
Shanghai Moule Technology (BUGBANK) started as a crowdsourced vulnerability testing platform. For this CyberGym test, the company participated alongside Geek Valley and the AI Security Industry Alliance (AISIA), entering with its self-developed security model and agent system. From an industry trend perspective, the focus of competition in the AI security field is shifting from “model parameter scale and tool stacking” toward “compliant deployment capabilities and real-world effectiveness on long-running tasks.” Building a continuously evolving vertical security system based on front-line practical experience is emerging as an engineering path worthy of attention.
About BUGBANK
BUGBANK is a prominent security brand under Shanghai Moule Network Technology Co., Ltd. As a world-leading platform for the discovery of cybersecurity vulnerabilities, it pioneered the concepts and defense frameworks centered on “open security” and “advanced threats.” BUGBANK collects, analyzes, responds to, and tracks the latest vulnerabilities emerging across the internet. It provides immediate alerts regarding zero-day vulnerabilities and critical cybersecurity threats, striving to establish itself as the premier brand in internet security.
Contact:
Sun Wengguo
18022471737
986883258@qq.com
https://www.bugbank.cn/
SOURCE: BUGBANK
View the original press release on ACCESS Newswire
Media gallery


