LLM EVALUTATIONS MISCONCEPTIONS:-

• Benchmarks ≠ Safety & Security
• Foundational Model ≠ LLM App

LLM APPLICATION SHARED RISKS:-

• Toxicity & Offensive Content
• Criminal & Illicit Activities
• Bias & Stereotypes
• Privacy & Data Security

LLM APPLICATION UNIQUE RISKS:-

• Inappropriate Content
• Out of Scope Behaviour
• Hallucinations
• Sensitive Information Disclosure
• Security Vulnerabilities

IDEAS & RESOURCES TO CHECK LLM APPLICATION SAFETY:-

1. OWASP Top 10 For LLM Applications
2. AIID - AI Incident Database
3. AVID - AI Vulnerability Database

SOME EXAMPLES OF MAJOR LLM VULNERABILITY CATEGORIES SHOWN BELOW:-

• Bias & Stereotypes