vmblog logo 2024 wht (updated)

Pentesters now use AI to build exploits, but only 2% trust it to chain attacks without supervision

Share: 

David Marshall | Published: November 17, 2026

Cobalt published its Autonomous Pentesting Pulse Report. The report is based on a survey of pentesters in the Cobalt Core, its community of 500+ trusted security experts.

Pentesters have moved their use of AI from writing reports to brainstorming attack paths and building exploits. Nearly 59% say exploit development and proof-of-concept creation has shifted significantly from manual effort to AI in the past six to 12 months. Brainstorming complex attack paths is pentesters’ top use of AI (49% of AI users, up 13 percentage points from 36% in January), while report writing fell as the most valuable use, from 55% to 39%. 

Adoption is also deepening. Ninety-eight percent of pentesters use AI, and those using it frequently or extensively nearly doubled since January, from 32% to 61%. But, trust has not kept pace, just 2% feel they could trust current AI models to chain a known exploit into a multi-step attack without human supervision. Furthermore, 37% now state human involvement is needed to validate automated scan results, up from 16% in January.

The report also found that:

  • 29% of pentesters agreed AI is a force multiplier that helps them find more high-impact bugs, up from 10% in January. 
  • Those calling AI a minor enhancement fell from 35% to 19%
  • Confidence in autonomous scanning is growing. The share ranking it ‘least effective’ fell from 64% to 26% since January.
  • Human-led PTaaS is still deemed the most effective model for finding complex, high-impact vulnerabilities, by nearly half (49%). This is more than double any other model, ahead of traditional time-boxed pentesting (18%), private bug bounty (15%), open / public bug bounty (12%) and autonomous / AI-driven scanning (6%).

“In the past nine months, the pentest community has rapidly evolved from using AI for mundane tasks to the point where it’s now integral to reconnaissance, reporting, and exploit development,” said Gunter Ollmann, CTO, Cobalt. “But it’s rightly not fully trusted. The share of pentesters who say human involvement is needed to validate results has more than doubled. Human oversight of all forms of AI remains critical, and those that let it run without it risk doing more harm than good.”