CyberGym AI bug hunting results put Microsoft and Wiz agents above 90%

Microsoft and Google-owned Wiz both reported CyberGym AI bug hunting scores above 90%, saying multi-model agent systems outperformed single-model approaches on vulnerability analysis. Wiz said its Project Atlas agent reached a 90.9% success rate and found more than 200 zero-day flaws in widely used open-source software, while Microsoft said its MDASH system scored 95.95% on the same benchmark.

The results put both vendors ahead of several frontier cyber models cited in the comparisons. Wiz said Atlas beat Anthropic’s Mythos Preview and OpenAI’s GPT-5.5 Cyber. Microsoft said MDASH outscored Mythos, Gemini and GPT systems on CyberGym, a benchmark used to test whether AI systems can identify real vulnerabilities in code.

What did Microsoft and Wiz score on CyberGym?

Microsoft reported the highest figure disclosed in the announcements, with MDASH at 95.95%. Wiz reported 90.9% for Project Atlas.

The comparison set included OpenAI’s GPT-5.5 Cyber at 85.6%, OpenAI’s GPT-5.6 Sol at 83.6%, Anthropic’s Mythos 5 at 83.8% for reproducing target vulnerabilities, and Google’s Gemini 3.5 Flash Cyber in CodeMender at 83.2%. All of those numbers should be read as benchmark results reported by the companies or cited by them, rather than evidence of how the systems perform across every production codebase.

Why are the agents using more than one model?

Both companies attributed their results to routing different security tasks to different models. Nir Ohfeld, Wiz’s head of vulnerability research, told The Register that Atlas currently uses Claude Opus 4.6 with GPT-5.5, and that Wiz is working to add Gemini after its recent collaboration with DeepMind on Gemini Flash Cyber.

Wiz said Atlas grew out of internal work on how frontier models can support advanced code scanning. The company has not made Atlas commercially available. Ohfeld and Wiz researcher Yuval Avrahami wrote that the company tests new models with an internal tool called Cyber Model Arena, which evaluates performance across threat modeling, hunting, validation and proof generation. Their point was that the strongest model for reasoning through an exploit chain may not be the best at triage, so Atlas assigns stages to whichever model performs best on that task.

Microsoft described MDASH as a combination of red-team agents, which find and simulate exploitable vulnerabilities and attack paths, and green-team agents, which remediate issues. The system uses MAI-Cyber-1-Flash, based on Microsoft AI’s internally developed MAI-Thinking-1 reasoning model, alongside GPT-5.4.

Microsoft said MAI-Cyber-1-Flash is intended to handle as much as 90% of the work. MDASH detects, patches and validates vulnerabilities, then sends the remaining more difficult tasks to GPT-5.4. Hayete Gallot, executive vice president of Microsoft Security, said the company achieved the results by placing an off-the-shelf model inside a multi-agent, multi-model harness.

Cost is part of the pitch

Microsoft and Wiz also framed model routing as a cost-control strategy, not only a performance tactic. Mustafa Suleyman, CEO of Microsoft AI, said combining Microsoft’s smaller in-house model with GPT-5.4 cuts customer costs by 50% while improving results.

Wiz made a similar argument from the code-scanning side. Ohfeld said frontier models keep expanding what security teams can do, but deep scans are costly and can become outdated as repositories change. His case for Atlas is that a scanner should be able to use the best available model for expert reasoning while continuing to work as stronger models arrive.

The broader signal for security vendors is clear enough: AI vulnerability scanning is moving from single-model demos toward orchestration systems that select models by task. The harder question, still unresolved by benchmark scores alone, is whether those systems can deliver reliable continuous coverage in customer environments without turning model cost and validation work into a new operational burden.

This story draws on original reporting from The Register.