Executive Summary
The recent leap in model capabilities for security has fundamentally changed what is possible in application security. AI is highly capable at uncovering complex flaws that evade traditional scanners, pentests, and even experienced human reviewers. For security teams, however, the question is no longer whether to use AI, but how to make it part of an application security program that continuously protects applications at enterprise scale, instead of producing one-off scans that become stale as soon as code changes.
Earlier this week, we shared results from Wiz Atlas, our AI vulnerability research system that achieved a 90.8% success rate on CyberGym and uncovered more than 200 previously unknown vulnerabilities. Atlas demonstrated what's now possible when frontier AI models are applied to vulnerability research. It also reinforced an important lesson: an effective enterprise AI application security program requires much more than a powerful model or a single scanner. It requires a system.
In the rest of this post, we’ll explore how we think about AI-powered application security as a layered system, why different layers require different capabilities, and how they work together to balance speed, cost, and depth. We’ll then explain how we’re applying these principles as we build Wiz’s solution for agentic application security.
4 Key Lessons from Running AI Against Enterprise Code
Atlas showed the great potential of intensive, frontier-model vulnerability research. Building and operating it also helped us identify key principles required to bring these capabilities into enterprise application security:
Deep scanning everywhere doesn't scale. Frontier model AI analysis can uncover complex vulnerabilities, but is also expensive and point-in-time. When developers and coding agents are producing changes continuously, a deep scan performed once a quarter cannot protect a codebase that changes every hour. Enterprises need broad, continuous AI scanning across their codebase as the foundation, while reserving targeted deep scans for selected applications, environments, and risks.
Depth should be driven by real application risk. A code scanner sees code, but it may not know whether that code is deployed, internet-exposed, reachable, or connected to sensitive data. Cloud and runtime context are needed to determine which vulnerability requires immediate attention - and which one warrants a deeper, more expensive investigation.
Different models —and engines—excel at different tasks. Our Cyber Model Arena shows that models have distinct strengths across cybersecurity domains. But a multi-model approach is not enough: different engines also bring different techniques, context, and design choices. Enterprise application security therefore requires a multi-engine, multi-model system that can apply the right specialized engine to each application and task, rather than betting on a single engine.
Building a harness is an ongoing investment. Models do not operate themselves, they require constant orchestration, tuning, benchmarking, and integration into development and security workflows. Maintaining this infrastructure is an ongoing investment that enterprise security teams should not have to take on themselves.
Introducing AI Cyber Model Arena: A Real-World Benchmark for AI Agents in Cybersecurity
Wiz Research’s AI Cyber Model Arena benchmarks offensive AI security on 257 real-world challenges (zero-days, CVEs, API/web, and cloud across AWS/Azure/GCP/K8s) demonstrating what AI models and agents can really do
Read more
Taken together, these lessons point to one conclusion: enterprise AI application security cannot be solved by a single model, or even a multi-model engine. It requires a managed, context-aware, multi-engine system that scans broadly and continuously, goes deep where application risk warrants it, and operationalizes findings across enterprise workflows.
The Enterprise Strategy: Scan Continuously Everywhere, Go Deep Where Risk Matters
A scalable application security system should apply different levels of analysis based on the risk and purpose of each scan. We think about this as three complementary layers where different scans are situated on different parts of the speed, depth, and cost curve. The point of a layered system is to give each scan the depth its risk justifies, and no more.
Tier 1: Baseline Deterministic Scanning
Baseline scanning provides fast, reliable, cost-effective coverage for known vulnerability classes using industry-standard benchmarks and deterministic signatures. These checks should run frequently, including during development and before code is merged, so entire classes of known risk are prevented rather than triaged later.
Tier 2: Continuous AI Reasoning
This scanning applies AI reasoning continuously across the application portfolio.
It is designed to identify vulnerabilities that deterministic pattern matching may miss, including risks that require understanding application structure, trust boundaries, data flows, and logic across multiple functions or files.
The goal is to bring AI-level analysis to every repository and every pull request as a form of ongoing coverage. Deterministic scanners already run on every change, but they can't see the business logic flaws that AI-generated code increasingly introduces. Catching that class of risk means AI scanning has to run continuously too, not just as a periodic deep pass, but on every PR.
Tier 3: Targeted Agentic Pentesting
Deep analysis is reserved for high-value applications, critical code or infrastructure, and findings that justify additional investment. These investigations use specialized engines, longer-running reasoning, and more intensive analysis to uncover subtle vulnerabilities and multi-step attack paths.
No single engine will be best at every type of investigation, and the leader changes with every frontier release. The deep layer should therefore support multiple first- and third-party specialized engines, selected and orchestrated according to the application and the security task.
Across all three layers, application, cloud, identity, exposure, and runtime context should determine where deeper analysis provides the greatest value. For example a SQL injection may look equally severe to a code scanner whether it sits in an internal sandbox or an internet-facing API. Only combined context reveals which is deployed and exposed, who owns the code, and which one actually deserves a researcher's time.
How Wiz Is Building This System Balancing Speed, Cost, and Depth
Wiz is applying this layered model across the full application lifecycle: from the code where vulnerabilities originate to the cloud and attack surface where they become exposed. Baseline, continuous, and deep scanning run across both code and the live application environment, with the Wiz Security Graph connecting them into a single system.
Crucially, the system is open at every layer. Wiz can apply its own engine, ingest signals from other solutions, and is designed to orchestrate third-party agentic scanning engines over time. Every finding, regardless of source, is correlated on the Wiz Security Graph, and flows through the same validation, remediation, and lifecycle management workflows AppSec teams already use. The key value is that organizations are never locked into a single scanner, and you benefit from the best available capability as the field advances.
Layer 1: Rule Based Scanning
Quick, cost-effective validation for baseline compliance and code hygiene.
Code — Wiz Rule-Based SAST (GA): Automatically blocks known patterns, insecure code, and hardcoded secrets before they are merged into your repository.
Cloud — Wiz ASM (GA): Continuously identifies exposed assets, weak credentials, misconfigurations, and other known attack-surface risks. By connecting those risks to applications and code ownership, Wiz helps teams understand both where risk is introduced and where it becomes exposed.
Layer 2: Continuous AI Scans
Broad, continuous analysis designed to catch complex, logic-based, and AI-discoverable risks across your application code.
Code — Wiz AI SAST (Private Preview): This brings AI-powered analysis to your first-party code, surfacing the complex business logic flaws that traditional pattern-matching misses. It works the way a security researcher would: first building an understanding of the application, its entry points, and trust boundaries. From there, it forms hypotheses about how an attacker could abuse the logic and reasons across functions and files to test them. It is productized from day one: no custom configuration, no harness to tune, just AI scanning running inside the AppSec workflows your team already uses.
Cloud — Wiz Red Agent (GA): Complements continuous ASM scans by autonomously testing live environments and applications using AI, helping teams understand which exposures can lead to real attack paths.
Layer 3: Targeted AI Pentest
For applications and risks that warrant deeper investigation, Wiz is building targeted agentic code analysis and application pentesting capabilities.
Atlas is being productized as Wiz’s first-party specialized deep-analysis engine, bringing the vulnerability research capabilities demonstrated by the Atlas system into a managed enterprise workflow. Deep analysis can then be directed toward selected high-value applications, critical code, and risks that justify a more intensive investigation.
And because the system is open, Wiz can also orchestrate specialized third-party engines where they add differentiated capability. For example, Google CodeMender, is an early example, contributing complementary analysis and remediation. Over time, Wiz will route across multiple first- and third-party engines for analysis, validation, and remediation, always applying the right one for the task rather than tying your program to a single scanner.
This is also where the managed model matters most. Keeping deep analysis at the frontier means continuously evaluating new models, harnesses, and agents as they ship—the work behind Cyber Model Arena—and folding the winners into the system. Wiz runs that evaluation and orchestration so your team consumes a system that stays near the frontier, rather than building and maintaining one to keep pace.
What this Means for Wiz Customers
AI has fundamentally changed what's possible in application security. The challenge is no longer finding vulnerabilities, it's knowing where AI should spend its attention.
That is the role of Wiz Code. Powered by the Wiz Security Graph, it connects code with runtime, identities, deployments, and attack surface to determine which applications need deterministic scanning, which need continuous AI reasoning, and which justify deep agentic investigation. Instead of treating every finding the same, Wiz applies the right level of analysis to the right risk and keeps the underlying engines current so you don't have to.
The result is an application security program that scales with modern software development. Known risks are prevented before they merge. Developers receive prioritized, actionable fixes. Security teams focus on the vulnerabilities that genuinely matter. And organizations gain the benefits of frontier AI without paying frontier AI costs across every repository and every pull request.
We believe this is the future of application security: not a single AI scanner, but an intelligent orchestration system that continuously applies the right analysis to the right application at the right time. That's the foundation of Wiz's layered approach to agentic application security.
Wiz Code customers can join the Wiz AI SAST Private Preview to evaluate broad, continuous AI scanning within their existing Wiz Code workflows and help shape the next phase of targeted agentic application security.
