top of page

NetSPI and Synack Find Human Testers Still Lead AI Pentesting Agents

12 minutes ago
1 min read

NetSPI and Synack found that AI agents and human penetration testers produced the strongest results together in a new study of web application security testing. The companies' 2027 Vision Report compared 22 agent configurations against a human baseline drawn from 425 assessments completed by 88 entry-level testers.


Across five applications, combining an agent with a human expert uncovered more issues than either approach alone. On one public application, each side found 23 of 26 known issue types, while the combined work found all 26. On another run, Synack's Sara AI Pentesting identified 62% of known issue types on its own and 88% when paired with experts, according to the report.


Human testers outperformed open-source agent configurations by an average of 23 percentage points across issue difficulty levels. The best model and harness combination detected 44% of known issues, compared with a 49% baseline for entry-level testers. The company cautioned that the comparison measures a particular set of tools and test applications, not every possible AI testing workflow.


The study also highlights the cost of noisy output. With source code access, one agent found eight of nine known issues in a proprietary application. In static testing, however, an agent produced 87 findings against 32 known issues, leaving 55 extra reports for human validation.


The research ran from May through August 2026 against three public and two proprietary web applications. Findings were scored against curated vulnerabilities and evidence requirements. For security teams considering AI-assisted pentesting, the results point to a role for agents in exploration and evidence gathering while experienced testers validate impact, resolve ambiguity and direct the work.

bottom of page