Safety & Alignment (2013-2023)

Techniques for ensuring AI systems are safe, aligned, and robust.

Safety & Evaluation

• Holistic Evaluation of Language Models - Comprehensive evaluation framework
• TruthfulQA: Measuring How Models Mimic Human Falsehoods - Truthfulness measurement
• Adversarial Examples Are Not Bugs, They Are Features - Adversarial robustness insights

Jailbreaks & Attacks

• Prompt Injection Attacks on Language Models - Injection attack taxonomy
• Universal Adversarial Triggers for Attacking and Analyzing NLP - Universal adversarial examples

---

Key Insight: Safety and alignment are critical for deploying AI systems responsibly in production.