On July 27th, Assaf Rappaport's Wiz introduced Project Atlas, a vulnerability research system that assigns different parts of a security investigation to specialized AI agents and models, according to Forbes. Wiz told Forbes that Atlas scored 90.9% on CyberGym, ahead of published results from Anthropic, OpenAI and Microsoft.
The announcement is an early look at what Rappaport and co-founders Ami Luttwak, Roy Reznik and Yinon Costica can build with Google behind them. The four founders previously built cloud security provider Adallom, sold it to Microsoft in 2015, and ran cloud security functions inside Microsoft before founding Wiz in 2020. Google completed its $32 billion cash acquisition of Wiz on March 11th, placing Wiz inside Google Cloud while retaining its brand and multicloud product strategy.
Project Atlas also extends a thesis the founders have pursued across two companies: security products become more useful when they assemble context around a risk instead of producing an isolated alert. Wiz applies that approach across code, cloud infrastructure and runtime activity. Atlas applies it within the vulnerability research process, dividing work among agents that can inspect code, challenge findings and select a model suited to each task.
The system is the product
Nir Ohfeld, who leads vulnerability research at Wiz, gave Forbes the clearest description of the architecture: "Atlas is not a model. It's a system that uses multiple models."
That distinction matters because model rankings can change with each release. A system built around interchangeable models can route difficult reasoning to a frontier model, use cheaper models for high-volume review, and send candidate vulnerabilities through separate validation stages. The surrounding software, prompts, tools, retry logic and proof generation can remain useful when a new underlying model arrives.
Wiz had already reached a similar conclusion through its own evaluations. In February, Wiz published an AI Cyber Model Arena with 257 security challenges. Wiz found that results depended on the combination of model and agent scaffold, and that no pairing led every category. Project Atlas turns that research observation into a working vulnerability pipeline.
According to information Wiz provided to Forbes, Atlas has discovered more than 200 previously unknown vulnerabilities in open-source projects. Wiz also said an early version found a GitHub.com vulnerability that produced a $100,000 bounty. Wiz has not publicly released enough detail to independently assess those claims, including the affected GitHub systems, disclosure record, Atlas model configuration or false-positive rate.
Atlas availability is also undefined. Wiz and Google had not published a product page, technical paper, public leaderboard entry or access program for Atlas alongside the Forbes report. That leaves open whether Atlas will become a Wiz product, a Google Cloud service, an internal research system or a limited tool for selected software maintainers.
What the benchmark proves
The reported 90.9% result would put Atlas ahead of several disclosed CyberGym scores. OpenAI reported an 85.6% result for GPT-5.5-Cyber when it expanded Daybreak in June. Anthropic's Mythos Preview research described a model that could find and exploit vulnerabilities across major operating systems and browsers. Microsoft reported an 88.45% CyberGym score for MDASH, its multi-model agentic scanning system, in May.
The OpenAI figure also shows why the comparison needs precise sourcing: Forbes listed GPT-5.5-Cyber at 85%, while OpenAI's own June 22nd announcement reported 85.6%.
CyberGym is substantial, with 1,507 vulnerabilities across 188 software projects. Its primary task gives an agent a vulnerability description and the relevant codebase, then asks it to produce a proof of concept that reproduces the flaw. A high score demonstrates code navigation, reasoning and exploit reproduction under that setup. It does not, by itself, establish that Atlas will find unknown vulnerabilities during a blind audit of an unfamiliar production system.
Microsoft's result provides an important comparison because MDASH also uses an ensemble rather than relying on one model. Microsoft says MDASH coordinates more than 100 specialized agents and found 16 vulnerabilities that were included in its May security updates. The parallel development at Wiz and Microsoft suggests the architecture is becoming a shared direction for AI security research rather than a Wiz-specific discovery.
Google gives Wiz a larger model bench
Wiz began developing Atlas after its December 2025 ZeroDay.cloud competition, Forbes reported. The event offered researchers cloud and open-source targets, and Wiz said participants earned $320,000 after uncovering critical vulnerabilities. That gave Ohfeld's research operation evidence that AI-assisted vulnerability work was moving from demonstrations into repeatable workflows.
The Google acquisition changed the resources available for that work. Forbes reported that Wiz collaborated with teams across Google, including DeepMind, while developing Atlas. Access to multiple model families, security researchers and Google's computing infrastructure lets Wiz test the system as an orchestrator instead of tying it to one flagship model.
That is the strategic value of Atlas for Rappaport's team. Wiz entered Google with a security graph that connects risks across development and production. Project Atlas gives Wiz a way to generate new vulnerability intelligence upstream, before those weaknesses appear as customer exposures. If Wiz can connect validated findings to the code, cloud and runtime context it already sells, Atlas could feed the broader Wiz platform rather than remain a standalone research project.
The remaining work sits after discovery. A vulnerability only reduces risk when researchers validate it, coordinate disclosure and land a tested patch. OpenAI has explicitly organized Daybreak around that remediation loop, while Microsoft designed MDASH to debate, deduplicate and prove findings before they reach engineers. Wiz has supplied strong discovery numbers for Atlas. The test for Rappaport and Ohfeld is whether the system can consistently turn those findings into fixes without creating another queue for maintainers to triage.