gshc2020.com

CyberGym is a benchmark that tests the ability of AI agents to discover vulnerabilities, so the fact that Project Atlas outperformed Mythos suggests it could be a more effective defensive tool than Anthropic’s unreleased frontier model.

According to Nir Ohfeld, head of vulnerability research at Wiz, told me in a video interview that one of the key differentiators of Atlas is that “unlike Mythos or Daybreak, Atlas is not a model. It’s a system that uses multiple models.” The tool uses a collaborative network of specialized, task-oriented AI agents, which select the best model for a given task.

The release comes less than two weeks after Hugging Face was breached by an autonomous agent, an incident which OpenAI later claimed occurred when GPT 5.6 Sol and an unknown prerelease model escaped a sandbox environment and hacked the open-source model provider.

Not One Model

When Anthropic announced Mythos back in April, the AI lab claimed the model had already found thousands of high-severity vulnerabilities in every major operating system and web browser. It also launched Project Glasswing as an “urgent attempt” to put these capabilities to use for defensive purposes, allowing third-party companies including AWS, Apple, Microsoft, Cisco and CrowdStrike to experiment with using the model.

MORE FOR YOU

Wiz has also been experimenting with AI to mitigate vulnerabilities in critical software. According to information provided by the company, Wiz first began working on Project Atlas after the company’s ZeroDay Cloud event, held at BlackHat Europe in December 2025. The team shared that during the event, they discovered a record-breaking number of CVEs affecting critical infrastructure, with many of them found using AI.

After discovering how much vulnerability research was shifting toward AI-assisted systems, the company began building Atlas, before joining Google to collaborate on what it refers to as “cutting-edge AI models for cyber defenders”. Since joining Google, Wiz has worked with multiple teams across the company, including DeepMind, to improve the tool.

Ohfeld notes that Wiz is using ChatGPT and Opus as part of the solution, but notes that Atlas scores above both in benchmarks because “we are using each model for its specific strengths." Not only has the approach performed well on CyberGym, but Wiz also claims that Atlas has found more than 200 previously unknown vulnerabilities in open-source projects, with fewer hallucinations and false positives than other alternatives.

He added that earlier this year, an early version of Atlas found “the most severe vulnerability ever found in GitHub.com,” helping to patch a vulnerability that would have allowed an attacker to run commands on GitHub servers and access customer repositories. This resulted in a bounty payout of $100,000, the largest bounty paid out to a researcher.

Inside Atlas

While powerful offensive models like Mythos and GPT-5.5 Cyber haven’t yet been widely released, with access gated behind Anthropic’s Project Glasswing and OpenAI’s Trusted Access for Cyber, these capabilities can’t be kept off the market forever. After all, equally strong models will come to market, as GLM-5.2 has shown.

As a result, defenders must prepare for a world where vulnerabilities are exploited at machine speed. Atlas represents an alternative approach to securing software against these fast-moving threats. “

Project Atlas is a meaningful demonstration of how quickly AI-driven vulnerability research is moving from experimental model capabilities toward operational security systems," Katie Norton, senior research manager of DevSecOps and software supply chain security at market intelligence provider IDC, told me via email, noting that frontier AI models can already reason about vulnerabilities.

However, Norton warns that “discovery is outpacing the industry’s ability to coordinate remediation and disclosure in the open source ecosystem,” adding that initiatives like Linux Foundation’s Akrites and the White House’s Gold Eagle reflect growing recognition that AI-generated vulnerability findings cannot simply be delivered by overburdened project maintainers.

“The industry will need shared mechanisms to validate and deduplicate findings, coordinate fixes confidentially, prioritize the vulnerabilities that present the greatest risk, and help patches reach downstream users before disclosure creates additional exposure,” Norton said.

Vulnerability Discovery Is Changing

Over the past year, Google has developed a number of vulnerability discovery tools, including CodeMender, a tool initially released as a research project in October 2025, which can automatically find and fix code vulnerabilities, and also Gemini 3.5 Flash Cyber, a lightweight model designed to find and patch vulnerabilities, announced in July.

Atlas appears to be Wiz and Google’s answer to a threat landscape that’s increasingly moving at machine speed. Ami Luttwak, cofounder and CTO of Wiz, told me in a video call that Atlas outperforming Mythos “shows you that for vulnerability scanning, it’s not just about the model, it’s also about the system and about how you use different models together.”

Luttwak adds that “there is urgent work that defenders need to do" following the Hugging Face Breach.“The same system two years ago would be considered secure, but the same system right now is not secure if you’re not running the latest AI model on your applications from the outside and on your code."

With 48,000 new CVEs discovered in 2025, defenders need to be prepared for threat actors to be able to exploit vulnerabilities at scale with AI. Tools like Mythos and Atlas could potentially provide an answer to this increase in scale if defenders can discover how to use them to patch exposures more efficiently.