AI has made finding vulnerabilities the easy part of offensive security. The hard part is knowing which vulnerabilities to prioritize first. With AI agents, security teams are testing at a scale no humans can match. They run continuously, probe every change, generate their own payloads, and chain small weaknesses into real attack paths. The result is often more vulnerabilities than organizations can sort through. Meanwhile, adversaries are using AI to find and exploit weaknesses faster and organizations are pentesting only about one-third of their attack surface on average, according to research Synack commissioned with Omdia.
The answer is a combination of security experts and agentic AI to efficiently move from validated findings to prioritized risk. Someone has to decide what is real, what is reachable, and what to fix first, and security teams say as much. Our research confirms this is the preferred approach: 64% of organizations name agent-led testing with human oversight as their ideal operating model, and 79% of security leaders won’t act on AI-generated findings alone.
At a high level, AI contributes breadth, speed, and endurance in security testing. Give an agent a task, a toolkit, and the right guardrails, and it covers ground no team could. Deterministic software handles what software does reliably, and agents work open-ended where the target demands it. People contribute judgment, creativity, and accountability. They decide what to test, explore where no known pattern exists, and validate what matters to the business.
Software development is already moving this way, with the most effective engineers running teams of agents. Offensive security is heading in the same direction, with testers who have agents at their disposal to scale their knowledge and their work. As AI agents take on more of the testing, people take on the tasks that provide direction and meaning.
Looking ahead, there are four roles that define where human judgment matters most, and they follow the life of a penetration test, from before an agent runs to after the findings are delivered.
- Setting scope and guardrails. Before an agent runs, someone has to decide what it may touch. That means naming the systems in play, the data that’s off limits, and the conditions that stop a test. The people in this role must understand the business. Scope also deserves a skeptical eye. Asset inventories are often incomplete, so the scope on paper can miss real attack surface, and experienced testers know to question it. With these guardrails in place, agents can work faster and more safely.
- Steering agents as they work. Agents excel when given a task. Open-ended exploration, where the goal is to go off, try things, and report back, is where experienced testers still lead. Experts direct the work as it unfolds, pulling an agent back before it exceeds its directive and pointing it toward the paths worth pursuing. The agents extend each tester’s reach, and the tester’s judgment keeps that reach pointed at what matters.
- Finding business logic flaws. Some of the most damaging flaws sit in what the business intended. Picture an illustrative checkout flow where discount codes stack, or an application that enforces a permission check on one route and forgets another. Every request looks legitimate, and no known pattern flags it. Finding these takes creativity and an understanding of how the business operates. It’s also where human testers still find the most critical vulnerabilities.
- Turning validated findings into prioritized risk. A validated finding tells you a weakness exists. It does not tell you what to fix first. Experts chain findings together, judge the real business impact, and recommend which fix closes the most attack paths. A person owns that priority call, stands behind it in front of executives and boards, and confirms with a retest that the fix worked. AI shortens every step of that work, and accountability for the outcome stays with a person.
What the combination makes possible
Bringing together the best of man and machine is the model we adopted when we first launched Synack. Now, 13 years later, it’s the reason why we’ve merged with NetSPI. Together, our two companies deliver one platform for offensive security across the entire attack surface, from pre-production code to live production, continuously or on demand.
Security researchers from the Synack Red Team and pentesters from NetSPI give agentic AI context, direction, and guardrails. AI gives those experts the speed, persistence, and reach to test more and go deeper. Because the experts also test what the AI does not find, they can see where it missed and why, and feed that learning back so the agents improve with every engagement.
Attackers are moving with AI. It’s time that defenders start testing the way adversaries do, with agents for reach and speed and experts for judgment. Expert offensive security is scarce, it is expensive, and most organizations receive far less of it than their attack surface warrants. Yet AI is how that expertise reaches more of the world.
As you consider your own security testing program and how it can combine the best of human experts and AI agents, consider these two questions: how much of your attack surface does it reach, continuously? And who stands behind what it tells you?
