How Security Teams Are Really Using Agentic AI
Security leaders aren’t waiting to see how agentic AI plays out. They’re already betting on it, and they’ve developed strong opinions about what separates a real penetration testing solution from a rebranded scanner or other DAST tools.
In fact, recent research from Fortune and Lightspeed Ventures shows that the vast majority of security teams are dedicating budget to AI—and 86% are dedicating 5% of their total budget or more. To dig deeper into how enterprise organizations are adopting AI specifically for pentesting, Synack commissioned a study with Omdia. Here’s what we found.
Trust in Agentic AI for Pentesting is Here
The report shows that 87% of organizations have moved beyond the evaluation phase and are actively planning, piloting, or using agentic AI for pentesting. More striking is the level of confidence behind this adoption: 87% of security leaders report high or complete trust in agentic AI’s ability to effectively test their enterprise environments. Among those who have fully deployed agentic AI for pentesting, that number is even higher—production-stage users are 2.2X more likely to express complete trust compared to those still in the pilot phase.
This aligns well with the 2026 National Cyber Strategy of America, which specifically calls out the importance of “adopting Al-powered cybersecurity solutions to defend federal networks and deter intrusions at scale.” This signals that the US Federal Government believes some of these AI-led solutions are ready for mainstream adoption.
Accuracy is the No.1 Requirement
For agentic AI to be trusted for production environments, security leaders have set a high bar. Over two-thirds—69%—of those surveyed said they require an accuracy level of at least 85% compared with manual testing. But accuracy in offensive security is more than just finding the same vulnerabilities as a human. It also means not surfacing false positives that waste triage time and erode trust.
This is where the design of the AI matters as much as its coverage. Traditional vulnerability scanners flag findings based on surface-level signals—a software version number, a configuration pattern—without confirming whether a vulnerability is actually exploitable in a given environment.
That gap between detection and certainty is often where analysts lose time and where decision-making breaks down. Sara Pentest addresses this directly through a coordinated team of AI agents that attempt exploits for in-scope vulnerabilities against the identified assets. Sara goes beyond risks—surfacing actual exploitable vulnerabilities.
Transparency in AI is Critical for Trust
Along with accuracy, 58% of organizations cite transparent AI decision-making as the top factor for increasing trust. And 93% say comprehensive guardrails are either important or critical for the safe and effective operation of agentic AI.
Synack has invested heavily in building industry-leading guardrails as detailed in our Agentic AI Pentesting Guardrails ebook. From strict scope enforcement to stay within approved IP ranges and applications, to destructive command blocklists (e.g. SQL DROP, SQL DELETE), Synack has designed this with customer trust as a top priority. The erosion of trust, stemming from the risk of permanently deleting important data or causing unexpected downtime, must be actively addressed and avoided.
Knowing the importance of transparency, we designed Sara Pentest specifically for explainability. Every potential finding is independently re-tested by a dedicated verification agent and a human before it enters the final report so that we keep our promise of only surfacing the vulnerabilities that matter. The result is a 2-3 day turnaround with findings that security leaders can actually explain to their stakeholders—and trust enough to act on.
Agent-Led with Human Oversight as the Gold Standard
Overall, 64% of organizations identify agent-led with human oversight as their preferred operational model—and this preference is strongest among those with the most hands-on experience. Organizations already using agentic AI for pentesting are 1.6X more likely to view human oversight as a permanent requirement for all testing scenarios, compared to those still in the pilot phase. We are increasingly seeing companies who run a fully autonomous pentesting solution begin to hire human pentesters. Synack’s agent-led, human oversight model combines the scale of AI with the precision of human expertise.
Final Thoughts
The data points in one direction: 95% of organizations anticipate that agentic AI will displace traditional pentesting services to some degree, and one in four expects to conduct pentesting exclusively through agentic AI within the next three years. What the research makes equally clear is that the path to that future runs through accuracy, transparency, and human oversight—not around them.
The organizations getting the most value from agentic AI aren’t the ones who deployed it and stepped back. They’re the ones who built a model where machine speed and human judgment operate together, each doing what it does best. That’s the gold standard. And the data suggests it’s already working.
Frequently Asked Questions
Where can I find the full research report?
View the complete report The 2026 State of Agentic AI in Pentesting.
Do security leaders actually trust AI to conduct pentesting?
Yes—87% of organizations report high or complete trust in agentic AI. That trust is built on two foundations: demonstrated accuracy and transparent decision-making. In fact, 58% percent of leaders cite transparency in how the AI reasons and executes as the top factor for increasing trust, and 93% say robust guardrails are important or critical for safe operation.
How long does an agentic pentest take?
Solutions like Sara Pentest can execute a test in hours, with human-validated results delivered in 2-3 days—significantly faster than traditional weeks-long assessments and designed to align with modern development and deployment cycles.
What is Sara Pentest?
Sara Pentest is a penetration test powered by agentic AI and built on Synack’s Penetration Testing as a Service (PTaaS) platform. Unlike generic vulnerability scanners, it uses a multi-agent methodology to emulate real-world attacker behavior—mapping the attack surface, planning and executing controlled exploits, and independently verifying every finding before it reaches the report. Exploitable vulnerabilities are then validated by security experts from the Synack Red Team, delivering high-confidence, actionable results at a scale traditional pentesting can’t achieve.
What is Sara Triage?
Sara Triage is an agentic AI solution that autonomously validates the output of existing vulnerability scanners, determining which findings are genuinely exploitable before your analysts ever see them. Traditional scanners flag vulnerabilities based on surface-level signals and can’t confirm actual exploitability—generating noise that obscures real risk. Sara Triage performs proof-based validation on each finding, surfaces only confirmed exploitable vulnerabilities, and integrates directly with tools like Jira, ServiceNow, Splunk, and Tenable. When deeper analysis is needed, findings can be escalated to the Synack Red Team for expert human review.
Why is human oversight considered a permanent requirement, not a transitional one?
The research shows that organizations with the most experience using agentic AI are more committed to human oversight, not less—1.6X more likely to view it as a permanent requirement compared to those still piloting. This reflects a mature understanding of what AI does well (scale, speed, breadth) and where human expertise remains essential (complex attack chains, business logic, contextual judgment). Agent-led with human oversight isn’t a halfway measure. It’s the model that delivers both machine speed and human-validated accuracy.
