CyberGym
Named by 14 kept items in this workspace.
12 items name it under tools and 2 under classification.key entities.
Search every workspace for CyberGym. It matches the name as written, so it can find other things that share it.
The items that name it 14
Cyber benchmark within the Artificial Analysis suite; Kimi K3.1 Pro rose from 58% to 95%
prior work on small-model cyber exploit capability
Predecessor benchmark to CyberGym-E2E; begins with a known vulnerability and focuses on exploit generation
Evaluation subset used to measure cyber capability differences between standard and abliterated GLM-5.3
tests agents on more than 1,500 realistic security tasks
Standard industry benchmark for finding vulnerabilities, used to evaluate autonomous vulnerability discovery
2025 evaluation environment/benchmark
Upstream benchmark/harness style; a third-party-deployed instance was abused as the rooted launchpad's code-execution harness
public benchmark for AI-driven vulnerability work; evaluates whether a system can produce a working proof of concept for a known vulnerability
Benchmark used to evaluate cybersecurity vulnerability-finding/fixing performance; 3.5 Flash Cyber reaches competitive performance at the frontier on it
Benchmark evaluating AI agents against hundreds of real-world software vulnerabilities, used to test 3.5 Flash Cyber
- Agent Capability Is a System Design Problem: Lessons From a 90% Improvement on CyberGym | depthfirst
Benchmark of roughly 1500 real vulnerabilities in which an agent must produce a Proof of Concept input that triggers a described bug; depthfirst evaluated on the ARVO portion (about 1370 examples) with a 50-example dev set and about 1320-example final test set
All technologies in this workspace.