Key Takeaways
- • Dream's autonomous cybersecurity research system Hero achieved a record 96.6% score on UC Berkeley's CyberGym benchmark, ranking first globally and surpassing all previously published results from frontier and specialized AI systems.
- • Hero's performance is powered by Hercules, Dream's purpose-built cybersecurity model developed from open-weight GLM-5.2 foundations and trained extensively on real-world vulnerability research data generated by Hero itself.
- • CyberGym evaluates AI systems against 1,500+ real-world security challenges from open-source projects, testing practical vulnerability research capabilities—code analysis, binary inspection, security tooling, crash tracing, and proof-of-concept exploit development—rather than theoretical knowledge alone.
- • Dream's 96.6% result reflects the complete autonomous research system's performance across specialized AI agents and integrated security tools, positioning the Tel Aviv-based sovereign AI company ahead of competitors in autonomous cybersecurity benchmarking.
Summary
Dream, a Tel Aviv-based sovereign AI company focused on governments and critical infrastructure, announced on September 4, 2026, that its autonomous cybersecurity research system Hero has achieved the highest score ever recorded on UC Berkeley's CyberGym benchmark with a 96.6% result, securing first place globally. This performance surpasses all previously published scores from both frontier AI models and cybersecurity-specialized systems, marking a significant advancement in autonomous vulnerability research capabilities.
CyberGym is a comprehensive evaluation framework that tests AI systems against more than 1,500 real-world security challenges extracted from open-source projects. The benchmark distinguishes itself from theoretical knowledge assessments by measuring practical capability—whether AI systems can perform the actual work of vulnerability research, including analyzing code and binaries, operating security tools, tracing software crashes, and developing functional proof-of-concept exploits. This focus on applied performance rather than theoretical knowledge makes CyberGym a rigorous standard for evaluating autonomous cybersecurity AI.
Dream's 96.6% score reflects the performance of Hero as a complete autonomous research system orchestrating multiple specialized AI agents and integrated security tools across complex investigations. At the core of this system is Hercules, Dream's specialized cybersecurity model, which handles the reasoning and analytical work behind vulnerability research. Hercules was developed using the open-weight GLM-5.2 model as its foundation, then extensively retrained on real-world vulnerability research data generated by Hero itself, allowing it to develop specialized expertise in the practical demands of cybersecurity work.
The benchmark result positions Dream competitively in an emerging market for AI-driven cybersecurity automation, with particular relevance to institutional investors evaluating sovereign AI capabilities and critical infrastructure security solutions. The achievement demonstrates measurable progress in autonomous systems capable of handling complex, specialized technical work that requires reasoning, tool integration, and practical security expertise—capabilities with significant commercial and strategic applications for government and enterprise security operations.
CyberGym is a comprehensive evaluation framework that tests AI systems against more than 1,500 real-world security challenges extracted from open-source projects. The benchmark distinguishes itself from theoretical knowledge assessments by measuring practical capability—whether AI systems can perform the actual work of vulnerability research, including analyzing code and binaries, operating security tools, tracing software crashes, and developing functional proof-of-concept exploits. This focus on applied performance rather than theoretical knowledge makes CyberGym a rigorous standard for evaluating autonomous cybersecurity AI.
Dream's 96.6% score reflects the performance of Hero as a complete autonomous research system orchestrating multiple specialized AI agents and integrated security tools across complex investigations. At the core of this system is Hercules, Dream's specialized cybersecurity model, which handles the reasoning and analytical work behind vulnerability research. Hercules was developed using the open-weight GLM-5.2 model as its foundation, then extensively retrained on real-world vulnerability research data generated by Hero itself, allowing it to develop specialized expertise in the practical demands of cybersecurity work.
The benchmark result positions Dream competitively in an emerging market for AI-driven cybersecurity automation, with particular relevance to institutional investors evaluating sovereign AI capabilities and critical infrastructure security solutions. The achievement demonstrates measurable progress in autonomous systems capable of handling complex, specialized technical work that requires reasoning, tool integration, and practical security expertise—capabilities with significant commercial and strategic applications for government and enterprise security operations.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.