Secure every application with flexible pentesting that combines expert-led testing, autonomous validation, and actionable results
Learn moreIdentify exploitable risk across external networks, internal infrastructure, and cloud environments
Learn moreExtend your offensive security program with specialized assessments, adversarial testing, code review, and program-level support
Learn moreWhere the line between human expertise and AI now sits, and what’s moving it.
Buyers who say agentic AI will be critical or very important to their offensive security within two years, [1] versus pentesters who would trust current AI to chain a known exploit into a multi-step attack without supervision.
Pentesters ranking autonomous scanning the least effective testing model, January vs. September 2026 (down 38 percentage points).
Pentesters who now use AI frequently or extensively, up from 32% in January (up 29 percentage points).
Buyers expect agentic AI to reshape offensive security within two years, yet only 7% run AI agents at full autonomy today. [2] And while pentesters doing the work already rely on AI heavily, they still won't let it run alone. That gap between what buyers forecast and what practitioners trust is where your next budget decision sits.
The Autonomous Pentesting Pulse Report pairs Omdia buyer research with two surveys of the Cobalt Core, our community of 500+ vetted pentesters, taken nine months apart. It shows where AI can safely take on work now, what it costs as testing goes deeper, and how to verify what an autonomous system actually tested.
Security and technology leaders: A clear view of where autonomous testing fits in your program and where human-led testing still has to carry the weight.
Budget owners: The cost curve behind AI-delivered testing, and the questions to ask any vendor before you fund it.
AppSec and security practitioners: First-hand data from surveys of working pentesters on what AI now does well and what they still do themselves.
The boundary between human-led and autonomous testing moved in nine months. These steps help you capture AI’s gains without overextending.
We asked Cobalt Core pentesters which parts of a pentest have shifted significantly from manual effort to AI.
The answer depends on the task: some work has largely moved to AI, while the tasks that call for context and judgment are still led by people.
Click or tap any bar in the chart to see how far AI has moved into that task and what still needs a human.
Share of pentesters who say each task has shifted significantly to AI, September 2026
Shifted to AI: share of pentesters saying each task has shifted significantly from manual effort to AI-assisted or AI-automated in the past 6 to 12 months (select all that apply). Source: Cobalt Core pentester surveys, September 2026 (n=94) and January 2026 (n=198). "Most valuable use" figures are pentesters who use AI (September n=92; January n=192). The figure for where pentesters still focus manual effort comes from an open-ended question, with answers grouped by theme (70 answered). [2]
Click or tap a bar in the chart to explore that task. Click or tap it again to clear your selection.
It's now so routine that pentesters count on it and move on to the harder work.
Where pentesters see the value
27% of pentesters who use AI name automated recon as AI's most valuable use, barely changed from 26% in January.
Download the report to see how the boundary between AI and human work moved between January and September 2026.
As automated output grows, so does the job of making sense of it.
What still needs a human
37% say human involvement is needed to validate, prioritize, or interpret automated scan results, up from 16% in January (up 21 pp).
Download the report to learn why reliability is the gate that keeps humans checking AI's work.
The lowest of any task: knowing how an application is meant to work, and how it can be abused, still takes human context.
What still needs a human
50% say humans are needed to identify logic flaws that automated tools typically miss. Asked where they still focus manual effort, 59% of those who answered named business logic.
Download the report to find out why reliability and cost keep the deepest testing human-led.
Nearly as many as for recon, but writing an exploit isn't the same as chaining exploits into an attack.
What still needs a human
Only 2% would let current AI chain a known exploit into a multi-step attack without supervision, and 53% say humans are needed to coordinate multi-step, chained attacks.
Download the report to learn what AI would need to do before pentesters trust it to chain exploits, and when we expect proof-of-concept code to become standard.
More than any other task, and it's no longer where pentesters see AI's biggest payoff.
Where pentesters see the value
Among pentesters who use AI, report writing as AI's most valuable use fell from 55% in January to 39% (down 16 pp).
Download the report to see what security buyers say matters most when they choose a testing partner.
See where the line between human-led and AI pentesting sits today, and what is moving it.
Download the reportThey are not the same thing. In this report, autonomous testing means AI that plans, executes, and reports on testing without a human tester directing each step. The report recommends asking every vendor what their term means and how they show what was tested.
1. Source – Omdia, Next-generation Offensive Security Strategies Give Defenders the AI Advantage, May 2026. Importance of agentic AI to bolstering offensive security posture in the next 24 months. Results are not an endorsement of Cobalt. Any reliance on these results is at the third party’s own risk.
2. Source – Omdia, Next-generation Offensive Security Strategies Give Defenders the AI Advantage, May 2026. Current level of autonomy of the AI agents used for offensive security tasks. Results are not an endorsement of Cobalt. Any reliance on these results is at the third party’s own risk.
Omdia disclaimerThe Omdia research, data and information referenced herein (the “Omdia Materials”) are the copyrighted property of TechTarget, Inc. and its subsidiaries or affiliates (together “Informa TechTarget”) or its third-party data providers and represent data, research, opinions, or viewpoints published by Informa TechTarget, and are not representations of fact.
The Omdia Materials reflect information and opinions from the original publication date and not from the date of this document. The information and opinions expressed in the Omdia Materials are subject to change without notice and Informa TechTarget does not have any duty or responsibility to update the Omdia Materials or this publication as a result.
Omdia Materials are delivered on an “as-is” and “as-available” basis. No representation or warranty, express or implied, is made as to the fairness, accuracy, completeness, or correctness of the information, opinions, and conclusions contained in Omdia Materials.
To the maximum extent permitted by law, Informa TechTarget and its affiliates, officers, directors, employees, agents, and third-party data providers disclaim any liability (including, without limitation, any liability arising from fault or negligence) as to the accuracy or completeness or use of the Omdia Materials. Informa TechTarget will not, under any circumstance whatsoever, be liable for any trading, investment, commercial, or other decisions based on or made in reliance of the Omdia Materials.