Security Leaders Turn to AI Pentesting as Attackers Deploy AI-Developed Zero-Days
Share
Security teams are increasingly adopting autonomous AI agents for penetration testing as attackers begin using artificial intelligence to develop and deploy zero-day exploits. The shift towards agentic security comes as traditional vulnerability management struggles to keep pace with the speed of modern exploitation.
The Google Threat Intelligence Group (GTIG) recently identified a threat actor using a zero-day exploit believed to have been developed with AI assistance. The exploit targeted a two-factor authentication (2FA) bypass rooted in a logic flaw, a type of vulnerability that conventional static analysis and fuzzing tools frequently overlook. While attackers are leveraging AI to find flaws, 87% of security leaders in the United States are now planning, piloting, or actively using AI-driven pentesting to close these defensive gaps.
Automated Attacks Outpace Traditional Patching Cycles
The urgency to adopt autonomous testing is driven by the shrinking window between vulnerability discovery and active exploitation. Data from the Mandiant M-Trends 2026 report indicates that exploitation often begins before a vendor can issue a patch. According to research from VulnCheck, 23% of vulnerabilities added to its exploited list in the first half of 2026 were targeted by attackers on or before the day of public disclosure.
Manual penetration testing, while thorough, is often too slow to provide continuous coverage. Models from CyCognito suggest that reconnaissance and analysis alone can consume 20 to 30 hours of a manual pentest that lasts up to 95 hours. AI agents can automate much of this reconnaissance, reasoning about the application’s structure to find complex flaws like the unauthenticated remote code execution (RCE) vulnerability tracked as CVE-2026-68982 in Apache Airflow.
Assessing Context and Production Safety
As organisations evaluate AI pentesting vendors, security experts warn that demos on controlled targets rarely reflect real-world performance. A key differentiator for agentic tools is their ability to handle logged-in workflows and maintain context across different user roles. Without deep context, an AI agent functions merely as a faster scanner rather than a reasoning security tool.
A case study from Reflectiz involving a major insurer highlighted the effectiveness of autonomous agents in identifying logic flaws. During an authorised assessment, an AI pentesting agent discovered that a one-time passcode (OTP) request system could be manipulated to send login codes to any email address. The agent proved the vulnerability by successfully logging in with a code sent to its own inbox, a finding that traditional automated tools had missed six months earlier.
However, the move toward autonomous testing requires strict guardrails. Effective AI pentesting must involve validated findings rather than a list of suspected issues. Experts recommend that organisations run “bake-off” tests on their own production or staging environments rather than relying on vendor-provided demos. This allows teams to verify if the agent can navigate complex authentication flows, chain multiple vulnerabilities together, and operate safely without disrupting production services.



