Automated Pentesting and BAS: How AI Systems Like XBOW Outpace Human Researchers in Vulnerability Discovery

The rapid evolution of infrastructure demands continuous security validation rather than annual manual pentests. Automated pentesting and Breach and Attack Simulation (BAS) tools now enable organizations to test defenses daily by safely replaying attacker behaviors.

BAS platforms such as Cymulate, SafeBreach, AttackIQ, and Picus focus on individual techniques mapped to the MITRE ATT&CK matrix. They verify whether antivirus, SIEM, or EDR solutions detect and block actions like privilege escalation or lateral movement without searching for new vulnerabilities.

Autopentest solutions go further by chaining multiple steps into realistic attack paths that start from an entry point and reach critical assets. This approach answers whether an attacker can achieve an unacceptable event, providing direct feedback into vulnerability management workflows.

In Russia, Positive Technologies launched PT Dephaze with its commercial version released in February 2025. Version 3.0, released on 31 October 2025, added expanded attack chains, detailed evidence collection, and machine learning for fully controlled internal testing. The product received the National Runet Award in the Information Security category in December 2025. Other local offerings include APT BEZDNA from CTRLHACK and continuous pentesting platforms from BI.ZONE.

Globally, mature platforms such as Pentera and Horizon3.ai NodeZero already serve large enterprises. The most striking development occurred in 2025 when AI systems surpassed human researchers. Autonomous system XBOW accumulated over 2,000 reputation points and submitted more than 1,000 vulnerability reports, including 54 critical findings, securing first place on HackerOne within 90 days. In one benchmark, XBOW completed tasks in 28 minutes that took a 20-year veteran pentester 40 hours.

Google’s Big Sleep project, a collaboration between DeepMind and Project Zero, discovered CVE-2025-6965, a critical memory vulnerability in SQLite affecting all versions prior to 3.50.2. The team predicted imminent exploitation and coordinated a patch release within 48 hours, marking the first documented case of an AI agent preventing real-world exploitation.

These capabilities are being consolidated under Gartner’s new Automated Exposure Validation category, which merges BAS, autopentest, and external attack surface management. The findings loop directly back into vulnerability management, raising the priority of previously missed attack paths and ineffective controls.