Microsoft and Wiz mind-meld agents catch more than 90% of bugs

Security

Secret to their success: Using the right model for the right security job

Two agentic bug-hunting systems from Microsoft and Google-owned Wiz show that when it comes to finding and remediating software vulnerabilities, at least two models’ minds work better than one - and Wiz tells us it’s adding a third.

Wiz on Monday said Project Atlas, its bug-hunting AI agent, bested Anthropic’s Mythos Preview and OpenAI’s GPT-5.5 Cyber with its vulnerability-analysis skills, achieving a 90.9 percent success rate on CyberGym, and uncovering more than 200 zero-day security holes in widely used open-source code.

Meanwhile, Microsoft boasted its MDASH bug-hunting harness scored a 95.95 percent success rate on CyberGym, also beating Mythos, Gemini and GPT on the same benchmark for evaluating how well AI systems find real vulnerabilities in the code.

For comparison, OpenAI’s GPT-5.5 Cyber scored 85.6 percent on CyberGym, and its 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.

The secret to both Atlas and MDASH’s success, according to the vendors, is that they use the right model for the right security job.

Atlas uses Claude Opus 4.6 with GPT-5.5, Nir Ohfeld, head of vulnerability research at Wiz, told The Register. “We're now working to incorporate Gemini, which is well timed given Wiz's recent work with DeepMind on Gemini Flash Cyber,” he added.

Microsoft’s MDASH - a combination of red-team agents that find and simulate real, exploitable vulnerabilities and attack paths, and green-team agents that remediate the issues - combines MAI-Cyber-1-Flash, based on Microsoft AI (MAI)’s internally developed MAI-Thinking-1 reasoning model, and GPT-5.4.

MAI-Cyber-1-Flash is designed to handle up to 90 percent of all tasks, with MDASH detecting, patching, and validating vulnerabilities before handing the remaining 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 on Monday.

Atlas isn’t commercially available yet - it’s used internally, and stems from Wiz’s efforts to understand how frontier models can be used for advanced code scanning. But it’s proof that “no single model is best at everything, and none stays state of the art for long,” Ohfeld and fellow Wiz kid Yuval Avrahami wrote in a Monday blog.

The cloud security biz evaluates every new model using its internal benchmarking tool, Cyber Model Arena, which scores each one on its success at completing various security-investigation tasks: threat modeling, hunting, validation, and proof generation.

“The results are rarely uniform: the model that reasons best through a complex exploit chain is often not the one that triages most precisely,” the duo wrote. “Atlas routes each stage to whichever model wins on that task.”

In addition to doing a better job of finding and fixing vulnerabilities, a multi-model system also saves customers’ money, according to Microsoft and Wiz.

Combining its much smaller, in-house model with GPT-5.4 halves customers’ costs, according to Mustafa Suleyman, CEO of Microsoft AI.

“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 on Monday.

And while “each new generation of models expands what is possible,” they are also expensive, Ohfeld told us. “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,” he said. 

In fact, the real question for code security shouldn’t be which model a scanner uses, Ohfeld added.

It’s this: “How does your system take advantage of the best model available today, continuously and economically, and what continues to work when a better one arrives,” he said. “That is the bet behind Atlas: 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.” ®