AI Security Testing — Most AI Attacks Are Noise, a Few Leave Craters

The report argues that AI security failures fall into two regimes: frequent, low-impact noise (e.g., jailbreaks and malformed prompts) and rare, high-impact events with power-law characteristics that can cause systemic damage. It emphasizes that while guardrails and monitoring manage the noise, organizations must prioritize architectural foresight, rigorous threat modeling, separation of trust domains, and holistic oversight of AI orchestration to prepare for the few but devastating failures that matter most.