![]()
MINNEAPOLIS, Nov. 19, 2026 (GLOBE NEWSWIRE) — NetSPI+Synack today released its 2027 Vision Report: Pentesting in the Age of AI Agents. The research set out to answer whether AI agents can pentest like humans. Although AI agents surfaced issues human testers missed, none of them delivered a consistent, human-equivalent penetration test on its own.
“Too much of the confidence in AI-only pentesting rests on demos and benchmarks run against public labs,” said Dr. Mark Kuhr, Chief Technology Officer at NetSPI+Synack. “We tested agents on a mix of public applications and ones they’d never seen, then scored them against trained human testers. While the agents caught vulnerabilities our experts rarely report, the research shows the strongest results come from experts paired with AI.”
Key findings from the report include:
- AI and experts, together, find more: On every application reviewed, pairing a human expert with an agent uncovered more issues than either alone. On one public application, experts and agents each found 23 of 26 known issue types; combined, they found all 26. On another public application run, Sara AI Pentesting found 62% of known issue types, however that rate rose to 88% when Sara was paired with experts. AI agents add the most immediate value by taking on repetitive exploration and evidence collection, while human testers direct the work, resolve ambiguity, validate business impact, and chain findings into real attack paths.
- Human testers lead on coverage, and the gap grows with difficulty: AI agents produced valid findings, but human testers outperformed open-source agent setups by an average of 23 percentage points across all difficulty levels. That lead grew from 22 points on easy issues to 30 points on moderate ones.
- Better models and harnesses help, but don’t close the gap: Even the best model tested, with optimal harness support, detected 44% of known issues, five percentage points below the entry-level pentester baseline of 49%. Harnesses focused models through tools, context, and a testing process, but switching harnesses didn’t give a model new pentesting abilities.
- More findings can mean more noise: Given source code access, one AI agent found eight of nine known issues on a proprietary application. But in static testing, AI reported 87 findings against 32 known issues, a 2.7x inflation that created 55 extra reports needing human validation.
“Security teams should view AI as another tool in the toolbox,” said Will Strei, VP of AI Transformation at NetSPI+Synack. “AI can automate work, but automation without the guidance of human testers is an incomplete solution. Insourcing pentesting rarely works, because few teams can staff the expertise needed to direct AI agents for optimal outcomes. Whether you want to achieve continuous testing or match the capabilities of AI-enabled attackers, experts remain a critical resource for managing agents and devising a complete solution.”
The report also offers recommendations for security leaders evaluating pentesting solutions in terms of cost, speed, noise, and thoroughness. Download the full 2027 Vision Report: https://www.netspi.com/resources/research/2027-vision-report-pentesting-in-the-age-of-ai-agents/.
Methodology
Between May and August 2026, NetSPI+Synack tested 22 AI agent configurations, spanning frontier and open-source models across multiple agent harnesses as well as Sara AI Pentesting, against five web applications. Two of the applications are proprietary, so their vulnerabilities were not publicly available. Each application had a curated list of known vulnerabilities with defined evidence requirements, and every agent finding was scored against that list. Results were compared with a human baseline of 425 instructor-reviewed assessments completed by 88 pentesters across seven classes of NetSPI+Synack University, a six-month training program for entry-level penetration testers.
About NetSPI+Synack
NetSPI+Synack is the leader in offensive security, combining agentic AI with elite human expertise to help organizations identify, validate, prioritize, and remediate exploitable risk across their entire attack surface. Its platform delivers expert-validated testing in depth, continuously, or on demand, from pre-production code to live production environments, so security teams can separate vulnerability noise from real risk, including the growing threat of AI-driven attacks. With global reach and more than 13 million hours of real-world offensive testing experience, NetSPI+Synack serves many of the world’s largest enterprises and government organizations with one ambition: nothing left to exploit. NetSPI+Synack is headquartered in Minneapolis, MN. Learn more at netspi.com, synack.com and follow us on LinkedIn.
Media Contact
Katy Nally
Sr Content Marketing Manager, NetSPI+Synack
cnally@synack.com

Media gallery
