Lyrie AI Takes World's No. 1 CyberGym Score at 99.2%, Beating Google and Anthropic Entries
In one test, Lyrie Agent reproduced 1,495 of 1,507 real-world software vulnerabilities and created working exploits. This gave it a reported score of 99.2%, the highest listed result on UC Berkeley's CyberGym Level 1 benchmark. It scored higher than the listed entries from Google, Anthropic, OpenAI and Huawei Cloud.
Dubai, United Arab Emirates, October 24, 2026 -- DUBAI, UAE, October 24, 2026 — Lyrie AI, the autonomous cybersecurity company behind Lyrie.ai, has reported the highest score on the CyberGym Level 1 leaderboard: 99.2%. The system reproduced 1,495 of 1,507 real-world software vulnerabilities and created working proofs of concept that could trigger those vulnerabilities. It completed one trial for each test case. The result is dated September 15, 2026, and was checked against the public leaderboard on October 23, 2026.
CyberGym was developed by researchers at UC Berkeley. It tests whether AI agents can reproduce real software vulnerabilities using working exploits. This means the agents must actually trigger the vulnerabilities, not just identify them. At Level 1, an agent receives a description of a vulnerability and a version of the software's code that has not been fixed. The test covers 188 real software projects and allows one attempt for each test case.
Lyrie Agent has the highest reported score on CyberGym Level 1: a 99.2% success rate at reproducing exploits.
According to the official leaderboard, Lyrie Agent scored higher than every other listed entry from major AI labs around the world. Huawei Cloud scored 95.4%, xAI scored 91.3%, Anthropic scored 91.0%, Google scored 86.3%, OpenAI scored 85.6%, Tencent scored 84.8%, and Meta scored 59.0%. No other entry listed in the public table scored higher than Lyrie's 99.2%.
"Reproducing a real vulnerability with a working proof of concept is the strongest evidence an autonomous security agent can provide," said Guy Sheetrit, founder and CEO of Lyrie AI. "It means that reasoning and execution come together at machine speed — and that is exactly what our commercial platform delivers for authorized security testing and fixing vulnerabilities."
The system combines inputs created by AI, automated fuzzing, and a repeated process in which a generator improves its work using feedback from program execution. It then checks that the resulting test program actually triggers the software fault. Lyrie's published technical report explains all 1,495 test cases that the system solved.
The company's commercial platform uses the same approach of checking results before accepting them. It includes OMEGA, an autonomous penetration-testing system that maps possible attack targets, finds ways an attacker could get in, tests exploits, and links vulnerabilities together into attack paths. It also includes Lyrie Remedy, which examines security findings, creates targeted patches, installs fixes, and checks that they work while keeping the original files intact. The wider platform also offers web application defense, endpoint protection, and security tools for AI-agent infrastructure.
Lyrie AI is inviting venture capital firms, strategic investors, enterprise technology partners, and potential buyers of the company to help it deliver products at a larger scale, reach more customers, and show results across a wider market. The full analysis, checked score table, and explanation of the benchmark are available at Lyrie Research.
About Lyrie AI
Lyrie AI (Lyrie.ai), developed by OTT Cybersecurity LLC, builds cybersecurity technology for automated penetration testing, vulnerability scanning, fixing security problems, defending web applications, and protecting endpoints. Its founder and inventor is Guy Sheetrit. Its headquarters are in Dubai, UAE.
Media Contact
Guy Sheetrit — Lyrie AI — [email protected] (investors) / [email protected] (partnerships) — lyrie.ai
Benchmark disclosure: Lyrie Agent ranked No. 1 by reported score among CyberGym Level 1 entries focused on agents that used a single trial, based on a check made on October 23, 2026. The leaderboard calls the displayed measurement the "Reference success rate." Individual teams run the tests and submit their results. Because the runs involve randomness, small differences in scores may not mean that one system is meaningfully better than another. Agent-focused and Model-focused entries are listed in separate categories. CyberGym tests how well systems reproduce vulnerabilities. This result does not prove how well Lyrie's commercial protection or vulnerability-fixing products perform.
Contact Info:
Name: Guy Sheetrit
Email: Send Email
Organization: Lyrie AI
Phone: +1 (424) 666-9521
Website: [https://lyrie.ai](https://lyrie.ai)
Release ID: 89207246
If you find any errors, concerns, or inconsistencies in this press release that need attention, or if you need help requesting its removal, please contact [email protected] as soon as possible. This email address is the authorized contact for these issues. Sending emails to several addresses does not necessarily make the process faster. The team aims to respond within 8 hours and will work to correct any problems or help with removal when needed. Providing accurate and reliable information remains its top priority.