Ai-safety

LLM-based applications face critical security challenges in form of prompt injections and jailbreaks. This article dives into the key architectural improvements underpinning ModernBERT, and demonstrates how to fine-tune it for discriminating malicious prompts. Our model closely approximates the performance of Claude 3.7 and Gemini Flash 2.0 on a mixed benchmark (NotInject, BIPIA, Wildguard-Benign, and PINT), while maintaining low latency (<40ms).

As reported by the FBI IC3, digital scams inflict devastating impacts on our society. In this project, developed for RDI Berkeley LLM-Agents Course (CS294/194-196), I built a multi-agent system with AutoGen to help users identify scam attempts. This agentic solution achieved higher accuracy compared to a prompt baseline (88.3% vs. 69.5%).