Artificial intelligence is becoming remarkably good at finding security flaws in software, but experts agree that relying on AI alone to protect your systems is a recipe for disaster. The real breakthrough isn't just faster vulnerability detection; it's learning to blend AI's speed and pattern recognition with human judgment, experience, and understanding of business context that machines simply cannot replicate.

Why AI Is Suddenly Finding Vulnerabilities That Humans Missed for Decades

In early April 2026, Anthropic's Frontier Red Team announced that the company's Claude Mythos Preview model had identified thousands of high and critical-severity vulnerabilities in major software systems. The discoveries included a 27-year-old bug in OpenBSD that could allow remote attackers to crash any machine running the operating system, web browser exploits that could let criminals read data from other domains like banking websites, and weaknesses in cryptography libraries that could enable hackers to decrypt communications or forge certificates.

What makes this remarkable is that Claude Mythos Preview wasn't explicitly trained to hunt for security flaws. The model's reasoning capabilities, combined with its ability to process vast amounts of code, simply made it exceptionally good at the task. Large language models, or LLMs, excel at what security experts call "finding the needle in a haystack." You can point an AI agent at a massive codebase, and it will systematically identify specific vulnerabilities with speed that would take human researchers weeks or months to match.

"You can point an AI agent at a large codebase, and they're very good at finding the needle in a haystack," said Jeremy Katz, vice president of code security at Sonar, a company that offers code verification solutions.

Jeremy Katz, Vice President of Code Security at Sonar

The speed advantage is particularly striking. AI models can reason about how data flows through code across different layers of abstraction, something that traditional rule-based security tools simply cannot do. This semantic understanding mirrors how human security researchers actually think about code, but at machine scale.

What Happens When AI Gets the Answer Wrong?

Here's where the story gets complicated. While AI excels at finding vulnerabilities, it also generates false positives at scale. An AI tool might incorrectly flag a bug as a critical security issue when it's actually harmless, or overstate the severity of a real flaw. For open-source software maintainers, who are often volunteers managing projects in their spare time, this creates a triage nightmare.

The volume of AI-generated reports is already overwhelming security teams. Researchers working with open-source maintainers report a "drastic uptick" in vulnerability reports, many of which are real bugs worth fixing but not actually security vulnerabilities. The distinction matters enormously, but distinguishing between the two requires human expertise and context that AI cannot provide.

Additionally, AI tools themselves can be attacked through techniques like prompt injection, where malicious inputs trick the model into behaving unexpectedly. Conversely, the same AI capabilities that find vulnerabilities can also be weaponized to exploit them. Claude Mythos Preview, for example, can chain together separate but related vulnerabilities to create step-by-step exploits that grant root access to the Linux kernel, the core of the operating system.

How to Build AI Security Defenses That Actually Work

The solution isn't to reject AI tools or to rely on them blindly. Instead, leading security organizations are implementing layered approaches that treat AI as a powerful assistant rather than a replacement for human judgment.

• Adversarial Self-Review: Tools like Claude Code Security and Google's CodeMender conduct what's called an adversarial self-review pass, meaning they challenge and critique their own results before presenting them to humans. This additional layer of scrutiny reduces false positives and builds checks and balances into the detection process.
• Human Verification as Standard Practice: Every finding from an AI tool must be checked and confirmed by human security professionals who understand the business logic behind the code. These tools produce probabilistic outputs, not final verdicts, and they cannot substitute for secure design reviews or penetration testing conducted by experienced security teams.
• Dynamic Threat Modeling and Red Teaming: Organizations are using dynamic threat modeling to evaluate likely threats to AI systems as they evolve, and red teaming to assess the safety and security of AI systems themselves. This includes continuous adversary emulation, where security teams simulate real-world attack scenarios to validate not just technology but also the people and processes defending it.

"These tools produce probabilistic outputs. They're not the final verdict. They cannot act as a substitute for your secure design reviews or penetration testing reviews. You still need somebody who understands the business logic behind your code and reviews that. And anytime AI gives us a finding, it goes through a verification process. There's always a human in the loop so we create these trust boundaries," explained Nayan Goel, a principal application-security engineer at the financial services company Upgrade.

Nayan Goel, Principal Application-Security Engineer at Upgrade

The "Hacker in the Loop" Model: Where AI Meets Human Expertise

One emerging approach gaining traction in enterprise security is what some firms call the "Hacker in the Loop" model. This framework explicitly rejects the idea that either pure automation or pure human effort can adequately defend modern systems. Instead, it combines AI-driven discovery and analysis with human-led judgment, campaign design, and detection validation.

Suzu Labs, a cybersecurity firm focused on secure AI adoption, recently acquired Emulated Criminals, a boutique firm specializing in adversary emulation and continuous red teaming. The acquisition created a new Continuous Adversarial Operations practice that executes named adversary operations against client environments at enterprise scale. The team, led by former U.S. Special Operations and offensive cyber experts, brings a "train how you fight" methodology to security validation.

"AI is cheapening the discovery side of offensive security, and that is fine with us. Discovery was never where the real work was. The real work is judgment, campaign design, and detection validation. That is what Adversarial Exposure Validation is actually about, and none of it scales through automation alone. It requires Hacker in the Loop, the human-led layer of AEV that automation cannot replicate," said Mike Bell, Founder and CEO of Suzu Labs.

Mike Bell, Founder and CEO of Suzu Labs

This model delivers ongoing, multi-vector attack simulations including phishing, ransomware scenarios, insider threats, and physical intrusion testing. The goal is to validate not just whether technology works in isolation, but how people and processes respond to realistic threats. AI handles the repetitive, pattern-matching work; humans handle the strategic thinking, context, and judgment.

What's Next: Closing the Gap Between Detection and Remediation

As AI gets better at identifying code-security weaknesses and accurately classifying their severity, security teams face a new challenge: closing the gap between detecting vulnerabilities and actually fixing them at scale. The workflow of finding a flaw is only half the battle; remediation, or fixing the vulnerability, follows predictable patterns that AI may eventually help automate as well.

The broader lesson is clear: the future of cybersecurity belongs to organizations that view AI not as a replacement for human expertise but as a force multiplier. Speed matters, but so does accuracy, context, and judgment. The most effective defenses will be those that let machines do what they do best, finding patterns and processing information at scale, while keeping humans in the loop for decisions that require understanding, experience, and accountability.