Agentic AI Offense Meets Defense: 17M Actions in Record-Breaking Live Attack

The rapid evolution of agentic AI is reshaping the cybersecurity battlefield. This exercise wasn't a theoretical proof-of-concept; it was a live, operational demonstration of fully autonomous AI agents in attack and defense. For AI researchers and developers, the numbers reveal the immense scale and precision that modern AI models can achieve—and the urgent need for governance frameworks that keep pace with technology.

On August 3, 2026, AI-native cybersecurity firm Armadin and agentic security operations provider TENEX.ai announced the completion of the largest controlled live AI cyberattack ever recorded. Conducted against a globally critical institution, the three-day exercise pitted Armadin's autonomous attacker swarm against TENEX.ai's agentic SOC platform, generating 17 million offensive actions, 38 validated attack paths, and 238 security findings. In parallel, TENEX.ai triaged 100% of 101,169 alerts in real time and reconstructed the entire attack across 231 billion raw events—a forensic task estimated to require 2,400 hours (four months) of manual analyst effort.

“

In parallel, TENEX.ai triaged 100% of 101,169 alerts in real time and reconstructed the entire attack across 231 billion raw events—a forensic task estimated to require 2,400 hours (four months) of manual analyst effort.

The engagement marks a pivotal moment in cybersecurity. Historically, organizations rely on periodic penetration tests and signature-based scanners, which provide only point-in-time snapshots. As adversaries increasingly adopt agentic AI to execute machine-speed attacks, defenders must evolve to continuous, autonomous validation and response. Armadin stated that its Hyperattacks against live environments are already standing operating practice, not a proof of concept. TENEX.ai similarly runs continuous agentic, human-led detection and response for every customer. When both operating models were run simultaneously, the exercise demonstrated that AI-driven offense and defense can operate at scales and velocities unattainable by human teams.

The numbers are staggering. Seventeen million offensive actions translate to an average of over 65 actions per second across the 72-hour window. The swarm autonomously discovered attack paths traversing the external perimeter, internal network, and web applications, simulating a sophisticated adversary without causing actual harm. On the defense side, 101,169 alerts—a volume that would overwhelm even the best-staffed SOC—were triaged in real time, with zero human analyst involvement. The reconstruction of 231 billion events provided a complete attack narrative, showing how agentic AI can turn months of manual investigation into a near-instantaneous capability.