The End of Human-Based Cyber Defense | Yossi Torati, A Security

Checklist based assessments fail against agentic attackers because natural language interpretation creates rare, exploitable exceptions, so defenses must be continuously challenged to reveal edge‑case behaviors and stay aligned with evolving agent tactics.

The cost of mounting advanced cyber attacks is falling toward zero because frontier AI democratizes nation‑state skills and automates continuous probing, so attackers shift from selective targeting to broad opportunistic scanning.

Moving from human-in-the-loop to human-on-the-loop scales safe remediation because systems execute fixes autonomously while humans supervise decisions, letting organizations act at machine speed without losing human oversight.

Natural language agentic systems are hard to constrain because instructions are open to interpretation and agents can justify rare bypasses, which makes static, deterministic rules unreliable for edge cases.

Operationalizing frontier models for realistic security testing requires an agentic harness because attackers act as unauthenticated, environment aware agents, so you need an orchestrator that simulates black‑box workflows safely and stays model‑agnostic.

A closed-loop stack of autonomous offense feeding autonomous defense is necessary because realistic, automated red teaming exposes multi-hop attack chains and that context lets defenses prioritize and apply safe mitigations in real time.

If security tooling produces massive false positives, teams get swamped triaging noise, which prevents timely remediation of real, business‑impacting attack paths.

Traditional patch cycles cannot outpace weaponized AI because the multi-step vulnerability lifecycle takes weeks or months while frontier models can find and weaponize flaws instantly.