Using Gemini 4 Argon to protect against cyber attackers

**TL;DR: Wiz researchers used Gemini 4 for vulnerability discovery and shadow A/B testing, detecting more than twice as many vulnerabilities as other baseline models while consuming fewer tokens; in the "Hack for Good" project with Google DeepMind, Gemini 4 traced a bypassable public file download feature all the way to a master key and escalated privileges to administrator, helping fix high-risk vulnerabilities on government, hospital, and critical infrastructure websites.** ## The Real Threat of Cyberattacks Any successful cyberattack can cause real harm. Highly sophisticated cyberattacks have threatened the lifelines of national security. In this situation, defenders need to find vulnerabilities faster and fix them faster. Wiz researchers' role is precisely to help security teams catch up—identifying critical vulnerabilities that were previously extremely difficult to find and helping organizations fix them in less than a day. ## Shadow A/B Testing Gemini 4 in Real Systems The Wiz team conducted shadow A/B testing of Gemini 4 in real systems. The results showed: - Gemini 4 could discover **more than twice as many vulnerabilities as other baseline models** - While using **fewer tokens** After seeing Gemini 4's outstanding performance on short-duration tasks, the team tested it on longer-duration tasks, achieving substantial improvements over other models. Gemini 4 could work tirelessly for hours, connecting complex vulnerabilities and signals to uncover critical risks that other models had previously failed to detect. This demonstrates a leap in frontier models' capabilities on tasks involving long contexts and multistep reasoning: security analysts often need to establish connections across vast amounts of alerts and logs, while models can run continuously and independently connect the clues. ## "Hack for Good": Using AI to Scan Critical Public Assets Recently, Wiz and Google DeepMind jointly announced the **"Hack for Good"** project. In this project, the team uses Gemini 4 to scan public websites belonging to governments, hospitals, and critical infrastructure, uncovering vulnerabilities and helping fix them. The project's goal is clear: put frontier AI capabilities on the defenders' side, covering systems whose compromise could have a major societal impact. ## Case Study: A Privilege Escalation Chain in Healthcare Software In a healthcare software application, Gemini discovered a **public file download feature** that could be bypassed to download internal files from the underlying system. Gemini 4 successfully identified a **configuration file** containing the application's **master key**. It then executed a complete attack chain: 1. **Discovery**: Identify the bypassable public file download feature 2. **Acquisition**: Download the configuration file containing the master key 3. **Privilege escalation**: Through read-only operations, elevate this master key to administrator privileges 4. **Impact validation**: Determine the actual scope of the impact This meant attackers could have gained access and taken almost complete control of these systems without the operations staff knowing. With the leap in frontier AI capabilities, the team fixed this vulnerability, protecting a large amount of sensitive information. ## A Turning Point for Defenders Having seen Gemini 4's actual cybersecurity capabilities, Wiz believes defenders can now discover and fix vulnerabilities faster than ever before, gaining an advantage in their contest with attackers. From detecting twice as many vulnerabilities in shadow A/B testing to analyzing the complete chain from download bypass to master-key privilege escalation in the "Hack for Good" project, Gemini 4 demonstrated its practical value in real defensive scenarios: faster discovery, more accurate correlation, and more complete reconstruction of attack paths. Source: Using Gemini 4 Argon to protect against Cyber attackers — Google DeepMind (https://www.youtube.com/watch?v=ftkic7Tdpcg)