The "defenders always lose" story in AI cybersecurity just hit a serious wobble. On September 3, 2026, Dream, the sovereign AI company built for governments and critical infrastructure, announced that its autonomous cybersecurity research system, Hero, scored an eye-watering 96.6% on UC Berkeley's CyberGym benchmark. That is not merely a win. It is the highest score ever recorded on the benchmark, and it comfortably outpaces the published results of the world's leading frontier AI models.
For an industry that has spent the past year hearing that AI gives attackers an unassailable edge, this is the kind of breakthrough defenders have been waiting for. The same AI revolution that supercharges nation-state hackers, Dream argues, can hand the defenders an even bigger advantage. Hero is the company's opening shot in that fight.
A Record 96.6%: The Highest Score Ever on CyberGym
CyberGym is the leading public, reproducible benchmark for AI-driven vulnerability research. Level 1 measures how well a system reproduces a known vulnerability and proves it with a working exploit. The test is brutally concrete, and it is exactly the kind of benchmark that separates marketing demos from real capability.
- 1,507 distinct tasks drawn from 188 real open-source projects
- Each task requires differential validation, meaning the AI must actually trigger the bug, not just describe it
- A single clean PASS@1 score determines a full round of validation, with no room for guesswork
Hero cleared the bar at 96.6%. To put the jump in perspective, the base model behind the system started at just 77% before fine-tuning and before the autonomous harness was layered on. Closing that nineteen-point gap is a massive leap in a single release cycle, and it pushed Hero past every published frontier model result on the board.
The Sovereignty Angle: Frontier Results With Zero Data Leaks
The most striking part of the breakthrough is not the number, but how Hero reaches it. Dream built the system to run entirely on-premises, inside the walls of the governments and critical infrastructure operators it serves. No API calls leave the network, no source code, firmware, or configuration crosses the boundary, and the entire stack is purpose-built for defensive security work.
At the heart of the system sits Hercules, a cybersecurity-specialized model post-trained from the GLM-5.2 base. Running in fp8 precision on local NVIDIA infrastructure, Hercules pairs with Hero's autonomous harness of scoped AI agents. A planner breaks each research mission into a phased task graph, an explorer maps entry points, trust boundaries, and data-flow paths, and multiple worker agents then hunt in parallel, tracing call chains from entry point to crash site.
Every finding passes through an independent review gate before it is trusted, checking for four things:
- A concrete path from attacker-controlled input to the vulnerable function
- Realistic triggering conditions that an attacker could actually meet
- A genuine security-boundary violation, not a theoretical nitpick
- Real novelty, so the team is not burning hours rediscovering known bugs
Findings that clear those filters get a dedicated execution environment and a working proof-of-concept that triggers the bug dynamically. Nothing is counted unless it can be shown end-to-end. That rigor is what makes the 96.6% score credible rather than cinematic.
Beyond the Benchmark: More Than 1,000 Real Zero-Days
Where Dream's story becomes genuinely exciting is in what the benchmark does not capture. CyberGym focuses on a narrow mission, a single named vulnerability with source code provided. In the real world, Hero is doing far messier work.
Over the past year, running at scale across open-source codebases, binaries, and firmware, Hero claims to have autonomously discovered and validated well over 1,000 previously unknown zero-day vulnerabilities. These include a critical unauthenticated remote code execution flaw in protocols common to critical-infrastructure networks (tracked as CVE-2026-32746), plus discoveries spanning protocol daemons, SCADA systems, embedded firmware, and widely deployed open-source libraries. Each one was confirmed with a working exploit produced by the system, not a static suspicion.
The toolchain behind those discoveries is formidable. It combines an RFC analyzer presented at Black Hat, a binary emulation framework for firmware and network-device research, code property graphs, format knowledge, and structured validation, all wrapped into a suite of more than forty proprietary security tools.
For Dream, the message is simple and confidently optimistic. As cyber-defense enters what the company calls the age of the infinite attacker, defenders can no longer afford to choose between capability and data sovereignty. With Hero and Hercules, they no longer have to. Frontier-competitive results, running fully on-premises, with the entire research stack retained inside the organization's control, is the kind of win that has governments around the world already putting the system to work.
The highest benchmark score ever recorded is a milestone worth celebrating. That it arrives wrapped in a sovereign, no-leak deployment model makes it a genuine turning point for how critical infrastructure can defend itself in the AI era.
Comments