What Enterprise Buyers Want From AI Pentesting, According to the Data

Enterprise buyers want AI pentesting that expands coverage and testing frequency while producing findings they can trust and act on. In Omdia’s research commissioned by Synack, 64% of respondents preferred agent-led testing with human oversight. That preference matters when security leaders decide what to fund and how to judge a pilot. Speed has value when the team can explain what was tested, verify the result and use it to reduce exposure.That preference matters when security leaders decide what to fund and how to judge a pilot. Speed has value when the team can explain what was tested, verify the result and use it to reduce exposure.

What Enterprise Buyers Want From AI Pentesting, According to the Data

Key Takeaways

  • Coverage is a priority: respondents said their organizations pentest an average of 32% of their attack surface.
  • Accuracy expectations need a defined measure and a fair comparison; the survey is not a product benchmark.
  • Transparency, guardrails and expert oversight belong in the operating model from the start.
  • Measure the pilot through validated coverage, testing cadence and remediation outcomes.

The 2026 State of Agentic AI in Pentesting surveyed 200 US security leaders and practitioners responsible for AI cybersecurity investment at enterprises with at least 1,000 employees. Omdia fielded the survey from December 11 to 22, 2025; findings about deployed adopters use a subset of 83 organizations.

I read these findings as a set of customer expectations. They help define what a useful pilot should demonstrate, while leaving each buyer responsible for testing those expectations against their own environment.

Why are security leaders adopting AI pentesting?

The coverage gap is substantial. Respondents estimated that an average of 68% of their attack surface was untested, leaving security leaders with limited evidence about large parts of their environment.

Adoption is moving, too: 87% had progressed beyond evaluation into planning a pilot, piloting or deployment. That figure includes organizations preparing to test the technology, so it should not be read as 87% already running it in production.

The commercial question is where additional testing would make the greatest difference. A buyer might start with applications changing frequently or systems that receive limited assessment, then agree which assets and attack paths the pilot must examine.

Keep the inventory, depth of testing and exclusions visible. Reaching an asset does not mean every role, workflow or vulnerability class on that asset has been assessed; this agentic AI pentesting explainer provides background on the approach.

How does AI pentesting differ from automated vulnerability scanning?

Vulnerability scanning typically identifies potential weaknesses through signatures, configurations and other checks. AI pentesting can add adaptive investigation and controlled exploitation attempts to establish what those weaknesses permit in a particular environment.

Capabilities overlap, and an AI label does not settle the distinction. Some scanning tools perform validation, while an agentic testing product may cover only particular vulnerability classes; examine the actual method and evidence.

For a buyer, the useful output connects a finding to an affected asset, the conditions required for exploitation and the observed result. Our exploitability validation guide explains that evidence in more detail.

How does AI pentesting compare with traditional manual pentesting?

Agentic AI can make repeatable testing easier to run across an agreed scope. Expert testers contribute direction, creativity and business context, including investigations of authorization boundaries, business logic and connected attack paths.

In its analysis accompanying the survey, Omdia flags limitations in cross-system attack chaining and the potential for higher false-positive rates in purely autonomous approaches. These are analyst observations, not a measured failure rate for every AI product.

A pilot should therefore examine how the methods work together. Ask experts to investigate a representative business workflow alongside the automated coverage, then document what each found and what remains untested; why human expertise remains necessary expands on those responsibilities.

Can AI pentesting reliably prove that a vulnerability is exploitable?

AI-driven testing can produce evidence of exploitability for vulnerabilities within its capabilities and authorized scope. Reliability depends on the assessment conditions, verification process and limits of the demonstrated result.

A successful exploit against one endpoint establishes something specific about that endpoint under those conditions. It does not prove that the whole application was tested or that every possible attack path was found.

Evidence should survive review by the people responsible for fixing the issue. Preserve the affected asset, access conditions, relevant actions and observed impact, with a clear distinction between what was demonstrated and what was inferred.

Expert review can challenge severity, investigate business consequences and identify gaps in coverage. It also supports decisions about what to fix first, as discussed in our article on human validation as a trust layer.

What should buyers measure in an AI pentesting pilot?

The findings suggest five decisions to agree before testing begins. These are practical recommendations drawn from the research, rather than a scoring model tested by Omdia.

1. How will we judge accuracy against a meaningful baseline?

In the survey, 69% required at least 85% accuracy compared with manual testing before considering agentic AI effective. This expresses a buyer expectation; it is not an achieved result for Sara or another product.

Define the measure before accepting a percentage. Compare equivalent assets, access and time windows, distinguish false positives from missed weaknesses, and have qualified testers review disagreements; a high count of confirmed findings alone says little about what was missed.

2. Can our team understand and review the testing record?

Transparency was selected by 58% as one of the factors that would increase trust in agentic pentesting systems. For buyers, that makes the testing record an important part of the evaluation.

For the pilot, request a reviewable record of scope, relevant actions and evidence supporting each conclusion. A remediation owner should be able to understand the issue and reproduce it safely without relying on an unexplained AI assertion.

3. How will we control execution and handle an exception?

The report found that 93% considered comprehensive guardrails important or critical. Translate that expectation into agreed boundaries, permitted actions, data-handling rules, stop conditions and named escalation owners.

Test how scope restrictions and intervention work before expanding access. Our agentic AI pentesting guardrails guide provides further reading for that discussion.

4. Where will experts direct and extend the work?

The 64% preference for agent-led testing with human oversight should prompt a staffing discussion. Identify who sets the scope, investigates complex workflows, reviews disputed findings and can intervene during execution.

Reserve expert time for investigating what automation may miss as well as validating its output. The earlier analysis of the human-oversight finding explores that preference; the pilot should make those responsibilities concrete.

5. What will count as useful progress in coverage and cadence?

Record the starting inventory and testing frequency, then agree the additional assets, depth and retest cadence the pilot should achieve. Track validated findings through remediation and verified closure, alongside the time the team spends reviewing them.

Use the same definitions before and after the pilot. Otherwise, a larger asset count or a more frequent run schedule can appear to show improvement without demonstrating deeper assessment or reduced exposure.

What do experienced adopters report?

Among the 83 deployed adopters, 87% agreed that agentic AI enabled testing of a higher percentage of their attack surface, and 91% agreed that it enabled more regular testing. These are self-reported outcomes, not independently measured performance gains.

Deployed users were also 1.6 times more likely than respondents still piloting to regard human oversight as a permanent requirement. This comparison describes different groups at one point in time; it does not establish that deployment caused individuals to change their views.

For me, the useful lesson is to evaluate scale and oversight together. A full-spectrum offensive security program needs testing that fits the software life cycle, with expert investigation of critical workflows and recurring validation as systems change.

If the decision includes developing your own tooling, our build-versus-buy guide addresses the associated operating responsibilities. Either route needs a clear owner for testing quality and remediation follow-through.

Read the full Omdia research report, then explore the Sara AI Pentesting interactive demo to see a testing workflow and prepare questions for your own assessment. Agree product scope and pilot success criteria with the team before drawing conclusions about fit.

Frequently Asked Questions

Learn how the Synack Platform can secure your organization