Best-of-N jailbreaking: defending against automated LLM attacks

### Executive summary: The report describes the "Best-of-N" prompt-injection (jailbreak) technique that automates many obfuscated prompt variations to bypass LLM safety filters, highlights real-world risks such as potential data exfiltration in healthcare and regulatory exposure, and recommends defenses including output-based safety evaluation, statistical anomaly detection, and adversarial red-teaming (e.g., Giskard Hub).