"We're now working to incorporate Gemini, which is well timed given Wiz's recent work with DeepMind on Gemini Flash Cyber," Nir Ohfeld, head of vulnerability research at Wiz, told The Register.
Benchmark scores: CyberGym and CodeMender results
Two vendor-built agentic systems — Wiz's Project Atlas and Microsoft's MDASH — reported double-digit gains over several leading single-model competitors on public benchmark suites. Wiz said Atlas achieved a 90.9 percent success rate on CyberGym and that its work uncovered more than 200 zero-day security holes in widely used open-source code. Microsoft reported that MDASH scored 95.95 percent on CyberGym.
For comparison, vendors reported these single-model scores on the same or similar evaluations: OpenAI’s GPT-5.5 Cyber scored 85.6 percent and GPT-5.6 Sol scored 83.6 percent; Anthropic’s Mythos 5 reproduced the target vulnerability on 83.8 percent of CyberGym challenges; and Google’s Gemini 3.5 Flash Cyber in CodeMender achieved an 83.2 percent success rate.
How Wiz’s Project Atlas is assembled
Wiz described Project Atlas as a multi-model routing system that uses different models for specific security-investigation tasks. Atlas currently pairs Claude Opus 4.6 with GPT-5.5 and, according to Nir Ohfeld, the team is integrating Gemini. The company evaluates new models with an internal tool called Cyber Model Arena, which scores models on threat modeling, hunting, validation, and proof generation.
Wiz’s researchers argue that “no single model is best at everything, and none stays state of the art for long,” and that “the results are rarely uniform: the model that reasons best through a complex exploit chain is often not the one that triages most precisely.” Their architecture, they wrote, “routes each stage to whichever model wins on that task.”
Microsoft’s MDASH: red teams, green teams, and model handoffs
Microsoft’s MDASH is a harness that pairs red-team agents — which find and simulate exploitable vulnerabilities and attack paths — with green-team agents that remediate issues and validate fixes. MDASH combines MAI-Cyber-1-Flash, based on Microsoft AI’s internally developed MAI-Thinking-1 reasoning model, with GPT-5.4.
According to Microsoft, MAI-Cyber-1-Flash is designed to handle up to 90 percent of tasks within the harness; MDASH detects, patches, and validates vulnerabilities before handing the remaining roughly 10 percent of more complex tasks to the larger GPT-5.4. “We were able to take an off-the-shelf model, within our harness, a multi-agent and multi-model implementation, and we achieved the best results you could have,” Hayete Gallot, executive vice president of Microsoft Security, said.
Costs, continuity, and the vendors’ operational case
Both vendors framed multi-model, multi-agent designs as not only technically stronger but economically advantageous. Microsoft AI CEO Mustafa Suleyman said combining a much smaller, in-house model with GPT-5.4 halves customers’ costs: “As the models hand off between each other, they are not just able to deliver better performance than all of the other models combined, they do so at 50 percent of the cost,” he said.
Wiz warned against treating frontier models as one-off scanners. “We have also learned that pointing a frontier model at a codebase once is not a sustainable security strategy: deep scans are expensive, their results become stale as code changes by the minute, and a point-in-time analysis cannot provide the continuous coverage organizations need across every repository,” Ohfeld said. He framed the Atlas bet as delivering “frontier-model depth where expert reasoning is required, an architecture that improves as models evolve, and rigorous validation so every finding arrives with evidence, not just a plausible answer.”
What this means for technologists, procurement leaders, and open-source maintainers
- Technologists and security teams: Expect architectures that route discrete stages of an investigation to the model that performs best at those tasks — threat modeling, exploit-chain reasoning, triage, or proof generation — rather than relying on one monolithic model.
- Procurement leaders and enterprises: The vendors are making an explicit cost argument. Microsoft claims that a hybrid harness can cut costs by half compared with running only larger, frontier models; procurement will need to weigh those savings against licensing, integration, and validation requirements.
- Open-source maintainers: Wiz’s disclosure that Atlas found more than 200 zero-day holes in widely used open-source code is a concrete signal that agentic scanners can surface real defects at scale — and that maintainers may see an acceleration of responsible disclosure and remediation demand.
Both systems are presented as proofs of concept for a broader operational lesson: route work to the model that wins at the task, validate findings with evidence, and design for continuous coverage as models evolve. Atlas remains an internal tool used to study frontier-model behavior and is not commercially available, while Microsoft has positioned MDASH as a harness that pairs small, task-focused models with larger generalists. The practical test ahead will be whether those multi-model harnesses sustain their reported gains in real-world, continuously changing codebases and whether vendors can operationalize the validation and cost claims at scale.
