Dream says cyber AI system sets record on Berkeley benchmark

Dream says its autonomous cybersecurity system Hero scored 96.6% on UC Berkeley’s CyberGym benchmark, the highest result ever recorded. The company says the result shows specialized AI can outperform larger general-purpose frontier models on real-world security research tasks.

Why it matters: - Dream says the result shows specialized AI systems can outperform larger general-purpose models on practical cybersecurity work. - The score lands as governments and critical infrastructure operators face faster, more AI-amplified cyber threats. - Dream is positioning Hero as a sovereign AI tool for security teams that need to keep sensitive work entirely on their own infrastructure.

What happened: - Dream announced that Hero, its autonomous cybersecurity research system, scored 96.6% on UC Berkeley’s CyberGym benchmark. - The score ranks Hero first globally on the benchmark, according to Dream. - Dream says the result beat previously published results from leading frontier and cybersecurity-focused AI systems. - The announcement was made in Tel Aviv, Israel, on September 4, 2026.

The details: - CyberGym tests AI systems on more than 1,500 real-world security challenges drawn from open-source projects. - The benchmark focuses on practical vulnerability research, including code analysis, binary analysis, security tooling, crash tracing, and proof-of-concept exploit development. - Dream says the 96.6% result reflects Hero as a complete autonomous research system, not the underlying model by itself. - Hero uses specialized AI agents and security tools to run complex investigations. - Hercules, Dream’s cybersecurity model, sits at the core of Hero and handles reasoning for code and binary analysis, dynamic debugging, data-flow tracing, crash investigation, exploit development, and remediation generation. - Dream built Hercules from the open weights of GLM-5.2. - Dream says it then trained Hercules further on real vulnerability research produced by Hero. - The training data came from Hero’s work across open-source codebases, binaries, and firmware. - Dream says that material went through automated and human review before being used to train Hercules. - The CyberGym evaluation ran entirely on Dream-controlled NVIDIA infrastructure. - Dream says the test environment had no public internet access and no external AI model APIs. - All data, model activity, and research stayed inside an environment fully controlled by Dream. - Shalev Hulio, Dream’s co-founder and CEO, said the result shows the AI race will not be decided only by the largest general-purpose model. - Hulio said specialized systems with the right tools can achieve extraordinary capabilities. - Hulio also said Hero can perform advanced cybersecurity work without internet access or external AI models while remaining under the user’s control. - Dream describes that approach as sovereign AI.

Between the lines: - The benchmark result is also a product message: Dream is trying to prove that security-specific systems can matter more than broad AI scale in narrow, high-stakes workflows. - By emphasizing on-premises operation, no internet access, and no external APIs, Dream is targeting buyers that care about control, isolation, and sensitive data handling. - Dream’s use of Hero-generated research to train Hercules suggests a feedback loop between autonomous execution and model improvement. - The company is also broadening the story beyond benchmark performance by pointing to reverse engineering, obfuscation bypass, forensic work, threat hunting, and remediation.

What’s next: - Dream says Hero is part of a broader sovereign AI portfolio for national defense. - The company says the broader platform is designed to help governments run advanced AI on their own infrastructure and data under their own control. - Dream is likely to use the CyberGym score to push deeper into government and critical infrastructure security programs. - More information is available on Dream’s website.

The bottom line: - Dream is betting that the future of cyber defense belongs to specialized, sovereign AI systems built for one mission, not general-purpose models built for everything.