The AI Pentesting wave. Pay for effort vs pay for impact.

The rise of AI in security testing has sparked debate about whether it will replace traditional bug bounty programs. According to Chris Holt, a security expert at Intigriti, the answer is a resounding no. The comparison between AI pentesting and bug bounty is fundamentally flawed, as they represent different approaches to security testing.

At the RSA Conference in March 2026, AI transitioned from a side topic to the default lens for security vendors. Every booth featured stories about AI agents, with messaging suggesting these tools replace human testers. Meanwhile, the security community had already begun experimenting with AI, producing initially messy outputs that improved as researchers learned proper tool usage.

Different Business Models

Bug bounty programs operate on a pay-for-impact model, where researchers are compensated only for validated vulnerabilities with sufficient context to fix them. This creates a direct incentive for quality over quantity. In contrast, most AI pentesting services follow a pay-for-effort model similar to traditional pentesting, but with fewer humans involved.

The AI pentesting model typically involves subscriptions and token consumption for continuous scans, producing streams of "possible issues" that require significant human triage. The real cost emerges after the initial scanning, as security teams must sort through false positives, duplicates, and determine exploitability.

The Three Waves of AI

When discussing the "AI pentesting wave," people are often referring to three distinct phenomena. In the community, researchers use AI to increase productivity without becoming obsolete. In the market, vendors are selling AI-first testing products that expand the security testing landscape. In boardrooms, executives hear "AI replaces humans" and push for budget shortcuts that don't match operational reality.

The AI Reality

AI follows a similar trajectory to previous security tools like vulnerability scanners, which raised the baseline but didn't eliminate pentesting. The danger with AI isn't its capability but the widening gap between findings and human capacity to validate them. "AI will find more bugs" misses the point when the pipeline can't absorb the output.

AI currently lacks depth in displaying impact, as it struggles with contextual understanding of boundaries, reachability, and exploitation paths. This leads to inflated findings that don't survive contact with reality. The most concerning hype is positioning AI as "autonomous" when achieving best results still requires significant tuning and validation.

The Future of Bug Bounty

In an AI-accelerated world, bug bounty programs must evolve with two parallel approaches: implementing intelligent friction to make low-effort submissions uneconomic, and using AI to scale the workflow through duplicate clustering, reproduction step extraction, and intelligent routing.

AI-first security tools excel at broad baseline coverage and pre-production pipeline automation, while bug bounty remains valuable for production systems, real attacker behavior, and complex environments. The key is maintaining the pay-for-outcomes core that aligns incentives with buyer needs: truth, exploitation reality, and reduced risk.

As the AI wave continues, responsible disclosure depends on honesty about incentives, economics, and bottlenecks. Bug bounty survives not by resisting change but by evolving to maintain its focus on validated vulnerabilities that matter in an era of abundant findings.