Autonomy is earned, not claimed

After more than 300,000 production penetration tests (pentests), our company has learned something that may surprise people watching the recent wave of autonomous security announcements.

The hardest problem in autonomous security isn’t teaching a machine how to attack. It’s teaching an AI-based system how to operate safely, predictably, and repeatedly inside production environments where mistakes have consequences.

Finding an attack path is an engineering problem. Building a platform that organizations trust to operate against healthcare systems, financial institutions, manufacturers, and critical infrastructure is an operational one. The difference only becomes apparent after years of running at scale.

As the industry embraces AI agents, autonomous red teaming, and machine-speed operations, much of the conversation remains focused on capability. Can a machine identify a path to compromise? Can it chain weaknesses together? Can it achieve the same outcome as a human operator?

Those are reasonable questions. They are not the questions security leaders ultimately care about.

Security leaders need confidence that a platform can operate safely in production, consistently produce meaningful results, and help teams make better decisions about risk. In our experience, that’s where the real challenge begins.

Since 2019, , a single compromised credential led to 586 critical impacts across 115 hosts, including three separate domain compromises. Viewed independently, the credential did not appear particularly significant. Viewed as part of an attack path, it became something entirely different.

In another , the larger question was how far an attacker could move after gaining access. Measuring blast radius exposed paths to systems and data that were never expected to be reachable from the original point of compromise.

These examples reinforce the same lesson. The challenge is rarely finding weaknesses. The challenge is knowing which weaknesses matter before an attacker does. That kind of judgment isn’t built from demonstrations or benchmarks. It’s earned through years of operating in production environments and seeing how real attack paths emerge across thousands of organizations.

Click